Fundemantals of Image Processing Course Project

• Ali Akay
• Mert Akkor

Hypothesis = Each individual Gan type leaves a unique noise residual on each genarated image and also PRNU(Photo-Response-Non-Uniformly) pattern. We would like to see our results under this hypotesis.

• In this project, our goal is to interpret compare the Gan's fingerprints between normal pictures residuals.
• We compared normal pictures with StyleGan and StyleGan2.
• For both StyleGans we used different psi values.

import pandas as pd
import numpy as np
import os
from glob import glob
from multiprocessing import cpu_count, Pool
import matplotlib.pyplot as plt
import numpy as np
from PIL import Image

import prnu

from multiprocessing import Pool, cpu_count

import numpy as np
import pywt
from numpy.fft import fft2, ifft2
from scipy.ndimage import filters
from sklearn.metrics import roc_curve, auc
from tqdm import tqdm

Generate Style GAN's Pics

• We used Flickr-Faces-HQ(FFHQ) dataset at 1024×1024 with styleGAN to generate our GAN pictures. The dataset consists of 70,000 high-quality PNG images at 1024×1024 resolution and contains considerable variation in terms of age, ethnicity and image background.
• We generated 300 GAN images like in the paper 'Does GANs leave artificial fingerprints?' by using styleGAN pre-trained model.

import tensorflow as tf
import os
import pickle
import numpy as np
import PIL.Image
import dnnlib
import dnnlib.tflib as tflib
import config

tf.compat.v1.Session()
dnnlib.tflib.init_tf()

WARNING:tensorflow:From C:\Users\user\Untitled Folder\prnu-python\stylegan\dnnlib\tflib\tfutil.py:109: The name tf.set_random_seed is deprecated. Please use tf.compat.v1.set_random_seed instead.

WARNING:tensorflow:From C:\Users\user\Untitled Folder\prnu-python\stylegan\dnnlib\tflib\tfutil.py:132: The name tf.ConfigProto is deprecated. Please use tf.compat.v1.ConfigProto instead.

#To generate images set your directory to stylegan folder

#pre traıned model
import pandas as pd
_G,_D,Gs = pd.read_pickle(r'karras2019stylegan-ffhq-1024x1024.pkl') 
# _G = Instantaneous snapshot of the generator. Mainly useful for resuming a previous training run.
# _D = Instantaneous snapshot of the discriminator. Mainly useful for resuming a previous training run.
# Gs = Long-term average of the generator. Yields higher-quality results than the instantaneous snapshot.

WARNING:tensorflow:From C:\Users\user\Untitled Folder\prnu-python\stylegan\dnnlib\tflib\network.py:142: The name tf.get_default_graph is deprecated. Please use tf.compat.v1.get_default_graph instead.

WARNING:tensorflow:From C:\Users\user\Untitled Folder\prnu-python\stylegan\dnnlib\tflib\network.py:150: The name tf.AUTO_REUSE is deprecated. Please use tf.compat.v1.AUTO_REUSE instead.

WARNING:tensorflow:From C:\Users\user\Untitled Folder\prnu-python\stylegan\dnnlib\tflib\tfutil.py:76: The name tf.VariableScope is deprecated. Please use tf.compat.v1.VariableScope instead.

WARNING:tensorflow:From C:\Users\user\Untitled Folder\prnu-python\stylegan\dnnlib\tflib\network.py:151: The name tf.get_variable_scope is deprecated. Please use tf.compat.v1.get_variable_scope instead.

WARNING:tensorflow:From C:\Users\user\Untitled Folder\prnu-python\stylegan\dnnlib\tflib\network.py:154: The name tf.placeholder is deprecated. Please use tf.compat.v1.placeholder instead.

WARNING:tensorflow:From C:\Users\user\Untitled Folder\prnu-python\stylegan\dnnlib\tflib\network.py:182: The name tf.global_variables is deprecated. Please use tf.compat.v1.global_variables instead.

WARNING:tensorflow:From C:\Users\user\Untitled Folder\prnu-python\stylegan\dnnlib\tflib\tfutil.py:200: The name tf.assign is deprecated. Please use tf.compat.v1.assign instead.

WARNING:tensorflow:From <string>:364: where (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.
Instructions for updating:
Use tf.where in 2.0, which has the same broadcast rule as np.where

Gs.print_layers()

Gs                              Params    OutputShape          WeightShape     
---                             ---       ---                  ---             
latents_in                      -         (?, 512)             -               
labels_in                       -         (?, 0)               -               
lod                             -         ()                   -               
dlatent_avg                     -         (512,)               -               
G_mapping/latents_in            -         (?, 512)             -               
G_mapping/labels_in             -         (?, 0)               -               
G_mapping/PixelNorm             -         (?, 512)             -               
G_mapping/Dense0                262656    (?, 512)             (512, 512)      
G_mapping/Dense1                262656    (?, 512)             (512, 512)      
G_mapping/Dense2                262656    (?, 512)             (512, 512)      
G_mapping/Dense3                262656    (?, 512)             (512, 512)      
G_mapping/Dense4                262656    (?, 512)             (512, 512)      
G_mapping/Dense5                262656    (?, 512)             (512, 512)      
G_mapping/Dense6                262656    (?, 512)             (512, 512)      
G_mapping/Dense7                262656    (?, 512)             (512, 512)      
G_mapping/Broadcast             -         (?, 18, 512)         -               
G_mapping/dlatents_out          -         (?, 18, 512)         -               
Truncation                      -         (?, 18, 512)         -               
G_synthesis/dlatents_in         -         (?, 18, 512)         -               
G_synthesis/4x4/Const           534528    (?, 512, 4, 4)       (512,)          
G_synthesis/4x4/Conv            2885632   (?, 512, 4, 4)       (3, 3, 512, 512)
G_synthesis/ToRGB_lod8          1539      (?, 3, 4, 4)         (1, 1, 512, 3)  
G_synthesis/8x8/Conv0_up        2885632   (?, 512, 8, 8)       (3, 3, 512, 512)
G_synthesis/8x8/Conv1           2885632   (?, 512, 8, 8)       (3, 3, 512, 512)
G_synthesis/ToRGB_lod7          1539      (?, 3, 8, 8)         (1, 1, 512, 3)  
G_synthesis/Upscale2D           -         (?, 3, 8, 8)         -               
G_synthesis/Grow_lod7           -         (?, 3, 8, 8)         -               
G_synthesis/16x16/Conv0_up      2885632   (?, 512, 16, 16)     (3, 3, 512, 512)
G_synthesis/16x16/Conv1         2885632   (?, 512, 16, 16)     (3, 3, 512, 512)
G_synthesis/ToRGB_lod6          1539      (?, 3, 16, 16)       (1, 1, 512, 3)  
G_synthesis/Upscale2D_1         -         (?, 3, 16, 16)       -               
G_synthesis/Grow_lod6           -         (?, 3, 16, 16)       -               
G_synthesis/32x32/Conv0_up      2885632   (?, 512, 32, 32)     (3, 3, 512, 512)
G_synthesis/32x32/Conv1         2885632   (?, 512, 32, 32)     (3, 3, 512, 512)
G_synthesis/ToRGB_lod5          1539      (?, 3, 32, 32)       (1, 1, 512, 3)  
G_synthesis/Upscale2D_2         -         (?, 3, 32, 32)       -               
G_synthesis/Grow_lod5           -         (?, 3, 32, 32)       -               
G_synthesis/64x64/Conv0_up      1442816   (?, 256, 64, 64)     (3, 3, 512, 256)
G_synthesis/64x64/Conv1         852992    (?, 256, 64, 64)     (3, 3, 256, 256)
G_synthesis/ToRGB_lod4          771       (?, 3, 64, 64)       (1, 1, 256, 3)  
G_synthesis/Upscale2D_3         -         (?, 3, 64, 64)       -               
G_synthesis/Grow_lod4           -         (?, 3, 64, 64)       -               
G_synthesis/128x128/Conv0_up    426496    (?, 128, 128, 128)   (3, 3, 256, 128)
G_synthesis/128x128/Conv1       279040    (?, 128, 128, 128)   (3, 3, 128, 128)
G_synthesis/ToRGB_lod3          387       (?, 3, 128, 128)     (1, 1, 128, 3)  
G_synthesis/Upscale2D_4         -         (?, 3, 128, 128)     -               
G_synthesis/Grow_lod3           -         (?, 3, 128, 128)     -               
G_synthesis/256x256/Conv0_up    139520    (?, 64, 256, 256)    (3, 3, 128, 64) 
G_synthesis/256x256/Conv1       102656    (?, 64, 256, 256)    (3, 3, 64, 64)  
G_synthesis/ToRGB_lod2          195       (?, 3, 256, 256)     (1, 1, 64, 3)   
G_synthesis/Upscale2D_5         -         (?, 3, 256, 256)     -               
G_synthesis/Grow_lod2           -         (?, 3, 256, 256)     -               
G_synthesis/512x512/Conv0_up    51328     (?, 32, 512, 512)    (3, 3, 64, 32)  
G_synthesis/512x512/Conv1       42112     (?, 32, 512, 512)    (3, 3, 32, 32)  
G_synthesis/ToRGB_lod1          99        (?, 3, 512, 512)     (1, 1, 32, 3)   
G_synthesis/Upscale2D_6         -         (?, 3, 512, 512)     -               
G_synthesis/Grow_lod1           -         (?, 3, 512, 512)     -               
G_synthesis/1024x1024/Conv0_up  21056     (?, 16, 1024, 1024)  (3, 3, 32, 16)  
G_synthesis/1024x1024/Conv1     18752     (?, 16, 1024, 1024)  (3, 3, 16, 16)  
G_synthesis/ToRGB_lod0          51        (?, 3, 1024, 1024)   (1, 1, 16, 3)   
G_synthesis/Upscale2D_7         -         (?, 3, 1024, 1024)   -               
G_synthesis/Grow_lod0           -         (?, 3, 1024, 1024)   -               
G_synthesis/images_out          -         (?, 3, 1024, 1024)   -               
G_synthesis/lod                 -         ()                   -               
G_synthesis/noise0              -         (1, 1, 4, 4)         -               
G_synthesis/noise1              -         (1, 1, 4, 4)         -               
G_synthesis/noise2              -         (1, 1, 8, 8)         -               
G_synthesis/noise3              -         (1, 1, 8, 8)         -               
G_synthesis/noise4              -         (1, 1, 16, 16)       -               
G_synthesis/noise5              -         (1, 1, 16, 16)       -               
G_synthesis/noise6              -         (1, 1, 32, 32)       -               
G_synthesis/noise7              -         (1, 1, 32, 32)       -               
G_synthesis/noise8              -         (1, 1, 64, 64)       -               
G_synthesis/noise9              -         (1, 1, 64, 64)       -               
G_synthesis/noise10             -         (1, 1, 128, 128)     -               
G_synthesis/noise11             -         (1, 1, 128, 128)     -               
G_synthesis/noise12             -         (1, 1, 256, 256)     -               
G_synthesis/noise13             -         (1, 1, 256, 256)     -               
G_synthesis/noise14             -         (1, 1, 512, 512)     -               
G_synthesis/noise15             -         (1, 1, 512, 512)     -               
G_synthesis/noise16             -         (1, 1, 1024, 1024)   -               
G_synthesis/noise17             -         (1, 1, 1024, 1024)   -               
images_out                      -         (?, 3, 1024, 1024)   -               
---                             ---       ---                  ---             
Total                           26219627                                       

def make_image():
    for i in range (500):
            # Pick latent vector.
            rnd = np.random.RandomState() 
            latents = rnd.randn(1, Gs.input_shape[1])

            # Generate image.
            fmt = dict(func=tflib.convert_images_to_uint8, nchw_to_nhwc=True)
            images = Gs.run(latents, None, truncation_psi=1.0, randomize_noise=True, output_transform=fmt)

            # Save image.
            os.makedirs(config.result_dir, exist_ok=True)
            png_filename = os.path.join(config.result_dir, 'example_'+str(i)+'.png'.format(1))
            PIL.Image.fromarray(images[0], 'RGB').save(png_filename)
            print(i)

make_image(500)

PRNU

from IPython.display import Image
Image(filename='prnu.png')

• We used generated 500 StyleGAN and StyleGan2 images, extracted their noise residuals and took the average of them. Then we compared it with a normal test image's noise residual by using normalized cross correlation.

"""ff_dirlist = GAN Images
    nat_dirlist = Natural Face Images
    """

ff_dirlist = np.array(sorted(glob(r'test\data\results\*.PNG')))
ff_device = np.array([os.path.split(i)[1].rsplit('_', 1)[0] for i in ff_dirlist])

nat_dirlist = np.array(sorted(glob(r'test\data\archieve2\*.PNG')))[:500]
nat_device = np.array([os.path.split(i)[1].split('0', 1)[0] for i in nat_dirlist])[:500]

stylegan_psi05 = np.array(sorted(glob(r'test\data\stylegan_psi0.5\*.PNG')))
stylegan_psi05_device = np.array([os.path.split(i)[1].rsplit('_', 1)[0] for i in stylegan_psi05])

stylegan_psi1 = np.array(sorted(glob(r'test\data\stylegan_psi1.0\*.PNG')))
stylegan_psi1_device = np.array([os.path.split(i)[1].rsplit('_', 1)[0] for i in stylegan_psi1])

stylegan2_psi05 = np.array(sorted(glob(r'test\data\stylegan2_psi0.5\*.PNG')))
stylegan2_psi05_device = np.array([os.path.split(i)[1].split('0', 1)[0] for i in stylegan2_psi05])[:500]
stylegan2_psi05=stylegan2_psi05[:500]

stylegan2_psi1 = np.array(sorted(glob(r'test\data\stylegan2_psi1.0\*.PNG')))
stylegan2_psi1_device = np.array([os.path.split(i)[1].split('0',1 )[0] for i in stylegan2_psi1])[:500]
stylegan2_psi1=stylegan2_psi1[:500]
    

stylegan_psi05.shape,stylegan_psi1.shape,stylegan2_psi05.shape,stylegan2_psi1.shape

((500,), (500,), (500,), (500,))

def threshold(wlet_coeff_energy_avg: np.ndarray, noise_var: float) -> np.ndarray:
    """
    Noise variance theshold as from Binghamton toolbox.
    :param wlet_coeff_energy_avg:
    :param noise_var:
    :return: noise variance threshold
    """
    res = wlet_coeff_energy_avg - noise_var
    return (res + np.abs(res)) / 2

def wiener_adaptive(x: np.ndarray, noise_var: float, **kwargs) -> np.ndarray:
    """
    WaveNoise as from Binghamton toolbox.
    Wiener adaptive flter aimed at extracting the noise component
    For each input pixel the average variance over a neighborhoods of different window sizes is first computed.
    The smaller average variance is taken into account when filtering according to Wiener.
    :param x: 2D matrix
    :param noise_var: Power spectral density of the noise we wish to extract (S)
    :param window_size_list: list of window sizes
    :return: wiener filtered version of input x
    """
    window_size_list = list(kwargs.pop('window_size_list', [3, 5, 7, 9]))

    energy = x ** 2

    avg_win_energy = np.zeros(x.shape + (len(window_size_list),))
    for window_idx, window_size in enumerate(window_size_list):
        avg_win_energy[:, :, window_idx] = filters.uniform_filter(energy,
                                                                  window_size,
                                                                  mode='constant')

    coef_var = threshold(avg_win_energy, noise_var)
    coef_var_min = np.min(coef_var, axis=2)

    x = x * noise_var / (coef_var_min + noise_var)

    return x

def pce(cc: np.ndarray, neigh_radius: int = 2) -> dict:
    """
    PCE position and value
    :param cc: as from crosscorr2d
    :param neigh_radius: radius around the peak to be ignored while computing floor energy
    :return: {'peak':(y,x), 'pce': peak to floor ratio, 'cc': cross-correlation value at peak position
    """
    assert (cc.ndim == 2)
    assert (isinstance(neigh_radius, int))

    out = dict()

    max_idx = np.argmax(cc.flatten())
    max_y, max_x = np.unravel_index(max_idx, cc.shape)

    peak_height = cc[max_y, max_x]

    cc_nopeaks = cc.copy()
    cc_nopeaks[max_y - neigh_radius:max_y + neigh_radius, max_x - neigh_radius:max_x + neigh_radius] = 0

    pce_energy = np.mean(cc_nopeaks.flatten() ** 2)

    out['peak'] = (max_y, max_x)
    out['pce'] = (peak_height ** 2) / pce_energy * np.sign(peak_height)
    out['cc'] = peak_height

    return out

def noise_extract(im: np.ndarray, levels: int = 4, sigma: float = 5) -> np.ndarray:
    """
    NoiseExtract as from Binghamton toolbox.
    :param im: grayscale or color image, np.uint8
    :param levels: number of wavelet decomposition levels
    :param sigma: estimated noise power
    :return: noise residual
    """

    assert (im.dtype == np.uint8)
    assert (im.ndim in [2, 3])

    im = im.astype(np.float32)

    noise_var = sigma ** 2

    if im.ndim == 2:
        im.shape += (1,)

    W = np.zeros(im.shape, np.float32)

    for ch in range(im.shape[2]):

        wlet = None
        while wlet is None and levels > 0:
            try:
                wlet = pywt.wavedec2(im[:, :, ch], 'db4', level=levels)
            except ValueError:
                levels -= 1
                wlet = None
        if wlet is None:
            raise ValueError('Impossible to compute Wavelet filtering for input size: {}'.format(im.shape))

        wlet_details = wlet[1:]

        wlet_details_filter = [None] * len(wlet_details)
        # Cycle over Wavelet levels 1:levels-1
        for wlet_level_idx, wlet_level in enumerate(wlet_details):
            # Cycle over H,V,D components
            level_coeff_filt = [None] * 3
            for wlet_coeff_idx, wlet_coeff in enumerate(wlet_level):
                level_coeff_filt[wlet_coeff_idx] = wiener_adaptive(wlet_coeff, noise_var)
            wlet_details_filter[wlet_level_idx] = tuple(level_coeff_filt)

        # Set filtered detail coefficients for Levels > 0 ---
        wlet[1:] = wlet_details_filter

        # Set to 0 all Level 0 approximation coefficients ---
        wlet[0][...] = 0

        # Invert wavelet transform ---
        wrec = pywt.waverec2(wlet, 'db4')
        try:
            W[:, :, ch] = wrec
        except ValueError:
            W = np.zeros(wrec.shape[:2] + (im.shape[2],), np.float32)
            W[:, :, ch] = wrec

    if W.shape[2] == 1:
        W.shape = W.shape[:2]

    W = W[:im.shape[0], :im.shape[1]]

    return W

import cv2

im = cv2.imread(ff_dirlist[0])
print(type(im))
# <class 'numpy.ndarray'>

print(im.shape)
print(type(im.shape))
# (225, 400, 3)
# <class 'tuple'>

<class 'numpy.ndarray'>
(1024, 1024, 3)
<class 'tuple'>

from PIL import Image

basewidth = 300
img = Image.open(stylegan_psi05[0])
wpercent = (basewidth/float(img.size[0]))
hsize = int((float(img.size[1])*float(wpercent)))
img = img.resize((basewidth,hsize), Image.ANTIALIAS)
img

from PIL import Image

basewidth = 300
img_2 = Image.open(stylegan2_psi05[1])
wpercent = (basewidth/float(img_2.size[0]))
hsize = int((float(img_2.size[1])*float(wpercent)))
img_2 = img_2.resize((basewidth,hsize), Image.ANTIALIAS)
img_2

from PIL import Image

basewidth = 300
img_3 = Image.open(stylegan2_psi1[1])
img_4 = Image.open(stylegan_psi1[1])
img_5 = Image.open(nat_dirlist[1])

plt.subplot(151)
plt.imshow(img, cmap=plt.cm.gray, vmin=40, vmax=220)
plt.axis('off')
plt.title('StyleGAN_psi0.5', fontsize=10)
plt.subplot(152)
plt.imshow(img_2, cmap=plt.cm.gray, vmin=40, vmax=220)
plt.axis('off')
plt.title('StyleGAN2_psi0.5', fontsize=10)
plt.subplot(153)
plt.imshow(img_3, cmap=plt.cm.gray, vmin=40, vmax=220)
plt.axis('off')
plt.title('StyleGAN2_psi1', fontsize=10)
plt.subplot(154)
plt.imshow(img_4, cmap=plt.cm.gray, vmin=40, vmax=220)
plt.axis('off')
plt.title('StyleGAN_psi1', fontsize=10)
plt.subplot(155)
plt.imshow(img_5, cmap=plt.cm.gray, vmin=40, vmax=220)
plt.axis('off')
plt.title('Natural', fontsize=10)

plt.subplots_adjust(wspace=0.02, hspace=0.02, top=0.9, bottom=0, left=0,
                    right=1)
plt.show()

def rgb2gray(im: np.ndarray) -> np.ndarray:
    """
    RGB to gray as from Binghamton toolbox.
    :param im: multidimensional array
    :return: grayscale version of input im
    """
    rgb2gray_vector = np.asarray([0.29893602, 0.58704307, 0.11402090]).astype(np.float32)
    rgb2gray_vector.shape = (3, 1)

    if im.ndim == 2:
        im_gray = np.copy(im)
    elif im.shape[2] == 1:
        im_gray = np.copy(im[:, :, 0])
    elif im.shape[2] == 3:
        w, h = im.shape[:2]
        im = np.reshape(im, (w * h, 3))
        im_gray = np.dot(im, rgb2gray_vector)
        im_gray.shape = (w, h)
    else:
        raise ValueError('Input image must have 1 or 3 channels')

    return im_gray.astype(np.float32)

rgb2gray(np.array(img)).shape

(300, 300)

d=noise_extract(np.array(img))
Image.fromarray(d, 'RGB')

c=noise_extract(np.array(img_2))
Image.fromarray(c, 'RGB')

def wiener_dft(im: np.ndarray, sigma: float) -> np.ndarray:
    """
    Adaptive Wiener filter applied to the 2D FFT of the image
    :param im: multidimensional array
    :param sigma: estimated noise power
    :return: filtered version of input im
    """
    noise_var = sigma ** 2
    h, w = im.shape

    im_noise_fft = fft2(im)
    im_noise_fft_mag = np.abs(im_noise_fft / (h * w) ** .5)

    im_noise_fft_mag_noise = wiener_adaptive(im_noise_fft_mag, noise_var)

    zeros_y, zeros_x = np.nonzero(im_noise_fft_mag == 0)

    im_noise_fft_mag[zeros_y, zeros_x] = 1
    im_noise_fft_mag_noise[zeros_y, zeros_x] = 0

    im_noise_fft_filt = im_noise_fft * im_noise_fft_mag_noise / im_noise_fft_mag
    im_noise_filt = np.real(ifft2(im_noise_fft_filt))

    return im_noise_filt.astype(np.float32)

def crosscorr_2d(k1: np.ndarray, k2: np.ndarray) -> np.ndarray:
    """
    PRNU 2D cross-correlation
    :param k1: 2D matrix of size (h1,w1)
    :param k2: 2D matrix of size (h2,w2)
    :return: 2D matrix of size (max(h1,h2),max(w1,w2))
    """
    assert (k1.ndim == 2)
    assert (k2.ndim == 2)

    max_height = max(k1.shape[0], k2.shape[0])
    max_width = max(k1.shape[1], k2.shape[1])

    k1 -= k1.flatten().mean()
    k2 -= k2.flatten().mean()

    k1 = np.pad(k1, [(0, max_height - k1.shape[0]), (0, max_width - k1.shape[1])], mode='constant', constant_values=0)
    k2 = np.pad(k2, [(0, max_height - k2.shape[0]), (0, max_width - k2.shape[1])], mode='constant', constant_values=0)

    k1_fft = fft2(k1, )
    k2_fft = fft2(np.rot90(k2, 2), )

    return np.real(ifft2(k1_fft * k2_fft)).astype(np.float32)

def extract_multiple_aligned(imgs: list, levels: int = 4, sigma: float = 5, processes: int = None,
                             batch_size=cpu_count(), tqdm_str: str = '') -> np.ndarray:
    """
    Extract PRNU from a list of images. Images are supposed to be the same size and properly oriented
    :param tqdm_str: tqdm description (see tqdm documentation)
    :param batch_size: number of parallel processed images
    :param processes: number of parallel processes
    :param imgs: list of images of size (H,W,Ch) and type np.uint8
    :param levels: number of wavelet decomposition levels
    :param sigma: estimated noise power
    :return: PRNU
    """
    assert (isinstance(imgs[0], np.ndarray))
    assert (imgs[0].ndim == 3)
    assert (imgs[0].dtype == np.uint8)

    h, w, ch = imgs[0].shape

    RPsum = np.zeros((h, w, ch), np.float32)
    NN = np.zeros((h, w, ch), np.float32)

    if processes is None or processes > 1:
        args_list = []
        for im in imgs:
            args_list += [(im, levels, sigma)]
        pool = Pool(processes=processes)

        for batch_idx0 in tqdm(np.arange(start=0, step=batch_size, stop=len(imgs)), disable=tqdm_str == '',
                               desc=(tqdm_str + ' (1/2)'), dynamic_ncols=True):
            nni = pool.map(inten_sat_compact, args_list[batch_idx0:batch_idx0 + batch_size])
            for ni in nni:
                NN += ni
            del nni

        for batch_idx0 in tqdm(np.arange(start=0, step=batch_size, stop=len(imgs)), disable=tqdm_str == '',
                               desc=(tqdm_str + ' (2/2)'), dynamic_ncols=True):
            wi_list = pool.map(noise_extract_compact, args_list[batch_idx0:batch_idx0 + batch_size])
            for wi in wi_list:
                RPsum += wi
            del wi_list

        pool.close()

    else:  # Single process
        for im in tqdm(imgs, disable=tqdm_str is None, desc=tqdm_str, dynamic_ncols=True):
            RPsum += noise_extract_compact((im, levels, sigma))
            NN += (inten_scale(im) * saturation(im)) ** 2

    K = RPsum / (NN + 1)
    K = rgb2gray(K)
    K = zero_mean_total(K)
    K = wiener_dft(K, K.std(ddof=1)).astype(np.float32)

    return K

Extract Fingerprints for Gan's

print('Computing fingerprints for StyleGAN psi=1.0')
fingerprint_stylegan_psi1 = sorted(np.unique(stylegan_psi1_device))
stg_ps1 = []
for device in fingerprint_stylegan_psi1:
    imgs = []
    for img_path in stylegan_psi1[stylegan_psi1_device == device]:
        im = Image.open(img_path)
        #im = im.resize((2000, 3008), Image.ANTIALIAS)
        im_arr = np.asarray(im)
        if im_arr.dtype != np.uint8:
            print('Error while reading image: {}'.format(img_path))
            continue
        if im_arr.ndim != 3:
            print('Image is not RGB: {}'.format(img_path))
            continue
        
        im_cut2 = prnu.cut_ctr(im_arr, (512, 512, 3))
        imgs += [im_cut2]
    stg_ps1 += [prnu.extract_multiple_aligned(imgs, processes=cpu_count())]
    
stg_ps1 = np.stack(stg_ps1, 0)

Computing fingerprints for StyleGAN psi=1.0

stg_ps1.shape

(1, 512, 512)

print('Computing fingerprints for StyleGAN2 psi=1.0')
fingerprint_stylegan2_psi1 = sorted(np.unique(stylegan2_psi1_device))
stg2_ps1 = []
for device in fingerprint_stylegan2_psi1:
    imgs = []
    for img_path in stylegan2_psi1[stylegan2_psi1_device == device]:
        im = Image.open(img_path)
        #im = im.resize((2000, 3008), Image.ANTIALIAS)
        im_arr = np.asarray(im)
        if im_arr.dtype != np.uint8:
            print('Error while reading image: {}'.format(img_path))
            continue
        if im_arr.ndim != 3:
            print('Image is not RGB: {}'.format(img_path))
            continue
        
        im_cut2 = prnu.cut_ctr(im_arr, (512, 512, 3))
        imgs += [im_cut2]
    stg2_ps1 += [prnu.extract_multiple_aligned(imgs, processes=cpu_count())]
    
stg2_ps1 = np.stack(stg2_ps1, 0)

Computing fingerprints for StyleGAN2 psi=1.0

stg2_ps1.shape

(1, 512, 512)

stg2_ps1

array([[[ 0.24823774,  0.02469793, -0.03474345, ...,  0.6418302 ,
         -0.11476105,  0.18489365],
        [-0.1565385 ,  0.0808139 ,  0.25029808, ..., -0.2669261 ,
         -0.39518213,  0.07902094],
        [ 0.18919197,  0.30362675, -0.32270437, ...,  0.24880326,
         -0.03878288,  0.08044937],
        ...,
        [-0.31729034, -0.34563464,  0.02844886, ..., -0.17093077,
          0.14621177, -0.00345003],
        [-0.2147267 , -0.28933984, -0.17443088, ..., -0.15213154,
         -0.1091008 , -0.26366845],
        [-0.4549669 , -0.42508656, -0.10889002, ..., -0.25044107,
         -0.11401205, -0.25126597]]], dtype=float32)

stg_ps1

array([[[ 0.40432537,  0.39162648, -0.25188863, ...,  0.6978138 ,
          0.5953538 ,  0.00078652],
        [ 0.61472356,  0.12943153,  0.10644206, ...,  0.0662146 ,
         -0.44741845, -0.41822878],
        [ 0.39647764,  0.4855263 ,  0.7520549 , ...,  0.30391923,
         -0.1270575 ,  0.29227784],
        ...,
        [ 0.12323871,  0.12040471, -0.23154517, ..., -0.10525158,
          0.13861366,  0.10875019],
        [ 0.37496755,  0.17548351,  0.2793322 , ..., -0.36677593,
         -0.0676221 ,  0.23318025],
        [ 0.14202394, -0.06342326,  0.08185127, ..., -0.510799  ,
         -0.22246107, -0.11330815]]], dtype=float32)

print('Computing fingerprints for StyleGAN psi=0.5')
fingerprint_stylegan_psi05 = sorted(np.unique(stylegan_psi05_device))
stg_ps05 = []
for device in fingerprint_stylegan_psi05:
    imgs = []
    for img_path in stylegan_psi05[stylegan_psi05_device == device]:
        im = Image.open(img_path)
        #im = im.resize((2000, 3008), Image.ANTIALIAS)
        im_arr = np.asarray(im)
        if im_arr.dtype != np.uint8:
            print('Error while reading image: {}'.format(img_path))
            continue
        if im_arr.ndim != 3:
            print('Image is not RGB: {}'.format(img_path))
            continue
        
        im_cut2 = prnu.cut_ctr(im_arr, (512, 512, 3))
        imgs += [im_cut2]
    stg_ps05 += [prnu.extract_multiple_aligned(imgs, processes=cpu_count())]
    
stg_ps05 = np.stack(stg_ps05, 0)

Computing fingerprints for StyleGAN psi=0.5

stg_ps05.shape
stg_ps05

array([[[ 0.6241586 , -0.22797832,  0.10719831, ...,  0.75395876,
          0.72783816,  1.3660004 ],
        [ 0.16785237, -1.6427523 ,  0.772056  , ..., -0.42739093,
         -0.08630094,  0.08583885],
        [ 0.4121152 ,  0.10297816,  1.045219  , ...,  0.93081903,
          0.39858156, -0.03737947],
        ...,
        [-0.38761693, -0.42904305,  0.06080198, ..., -0.0319345 ,
         -0.01931673, -0.09963929],
        [-0.326597  , -0.19424908,  0.5021167 , ..., -0.12286449,
         -0.3192697 , -0.05525027],
        [-0.47375783, -0.7964915 , -0.22155897, ..., -0.49982098,
         -0.22038022,  0.03944532]]], dtype=float32)

print('Computing fingerprints for StyleGAN2 psi=0.5')
fingerprint_stylegan2_psi05 = sorted(np.unique(stylegan2_psi05_device))
stg2_ps05 = []
for device in fingerprint_stylegan2_psi05:
    imgs = []
    for img_path in stylegan2_psi05[stylegan2_psi05_device[:500] == device]:
        im = Image.open(img_path)
        #im = im.resize((2000, 3008), Image.ANTIALIAS)
        im_arr = np.asarray(im)
        if im_arr.dtype != np.uint8:
            print('Error while reading image: {}'.format(img_path))
            continue
        if im_arr.ndim != 3:
            print('Image is not RGB: {}'.format(img_path))
            continue
        
        im_cut2 = prnu.cut_ctr(im_arr, (512, 512, 3))
        imgs += [im_cut2]
    stg2_ps05 += [prnu.extract_multiple_aligned(imgs, processes=cpu_count())]
    
stg2_ps05 = np.stack(stg2_ps05, 0)

Computing fingerprints for StyleGAN2 psi=0.5

stg2_ps05.shape

(1, 512, 512)

• Here, we obtained noise residuals from StyleGan and StyleGan2 with psi=0.5 and psi=1.0 each. Their average by using the function 'extract_multiple_aligned'. Later on we created their fingerprints and going to use them for comparison between normal pictures' residual.

Extract Fingerprints for Normal's

print('Computing fingerprints')
normal_device = sorted(np.unique(nat_device))
c = []
for device in normal_device:
    imgss = []
    for img_path in nat_dirlist[nat_device == device]:
        im = Image.open(img_path)
        #im = im.resize((2000, 3008), Image.ANTIALIAS)
        im_arr = np.asarray(im)
        if im_arr.dtype != np.uint8:
            print('Error while reading image: {}'.format(img_path))
            continue
        if im_arr.ndim != 3:
            print('Image is not RGB: {}'.format(img_path))
            continue
        
        im_cut1 = prnu.cut_ctr(im_arr, (512, 512, 3))
        imgss += [im_cut1]
    c += [prnu.extract_multiple_aligned(imgss, processes=cpu_count())]

c = np.stack(c, 0)

Computing fingerprints

• Here we did the same process like GAN pictures but for normal images to compare with normal pictures to understand relationship between normal fingerprints and residuals.
• We expriment it that normal-normal picture normalized cross correlation is higher than GAN-normal correlation, which lead us that we are in the right way.

c.shape

(1, 512, 512)

Computing Residuals

print('Computing residuals')

imgs = []
for img_path in nat_dirlist:
    imgs += [prnu.cut_ctr(np.asarray(Image.open(img_path)), (512, 512, 3))]

pool = Pool(cpu_count())
w = pool.map(prnu.extract_single, imgs)
pool.close()

w = np.stack(w, 0)

Computing residuals

w.shape

(500, 512, 512)

print('Computing residuals')
#stylegan_psi1_w=[]
imgs = []
for img_path in stylegan_psi1:
    imgs += [prnu.cut_ctr(np.asarray(Image.open(img_path)), (512, 512, 3))]

pool = Pool(cpu_count())
stylegan_psi1_w = pool.map(prnu.extract_single, imgs)
pool.close()

stylegan_psi1_w = np.stack(stylegan_psi1_w, 0)

Computing residuals

print('Computing residuals')

imgs = []
for img_path in stylegan2_psi1:
    imgs += [prnu.cut_ctr(np.asarray(Image.open(img_path)), (512, 512, 3))]

pool = Pool(cpu_count())
stylegan2_psi1_w = pool.map(prnu.extract_single, imgs)
pool.close()

stylegan2_psi1_w = np.stack(stylegan2_psi1_w, 0)

Computing residuals

stylegan2_psi1_w.shape

(500, 512, 512)

print('Computing residuals')

imgs = []
for img_path in stylegan_psi05:
    imgs += [prnu.cut_ctr(np.asarray(Image.open(img_path)), (512, 512, 3))]

pool = Pool(cpu_count())
stylegan_psi05_w = pool.map(prnu.extract_single, imgs)
pool.close()
stylegan_psi05_w = np.stack(stylegan_psi05_w, 0)

Computing residuals

print('Computing residuals')

imgs = []
for img_path in stylegan2_psi05:
    imgs += [prnu.cut_ctr(np.asarray(Image.open(img_path)), (512, 512, 3))]

pool = Pool(cpu_count())
stylegan2_psi05_w = pool.map(prnu.extract_single, imgs)
pool.close()

stylegan2_psi05_w = np.stack(stylegan2_psi05_w, 0)

Computing residuals

• We compute the residuals for 500 normal images in the above function. However, we just used 1 image as a test image.

stylegan2_psi05_w

array([[[ 2.04616189e-01, -1.01863921e+00, -1.43706226e+00, ...,
          1.98888421e+00, -4.84552622e+00, -7.66749680e-01],
        [ 6.33472741e-01, -2.06804013e+00, -2.87042427e+00, ...,
          4.47351408e+00, -2.58911937e-01, -1.86816227e+00],
        [ 2.13606071e+00, -4.17354870e+00, -2.11144710e+00, ...,
         -4.53416491e+00,  4.08587551e+00, -4.04173613e+00],
        ...,
        [ 1.15429866e+00, -6.91441357e-01,  1.14656538e-01, ...,
         -1.34751067e-01, -1.35562968e+00, -9.05156016e-01],
        [ 3.82139653e-01, -1.42756689e+00, -1.47295368e+00, ...,
          5.53542256e-01,  1.21271580e-01, -2.99411893e-01],
        [-1.01359499e+00, -6.74999416e-01, -2.80121708e+00, ...,
          1.20783520e+00,  1.62669623e+00,  1.68624997e-01]],

       [[-4.62847650e-01, -2.57518935e+00,  3.77967763e+00, ...,
         -2.33853515e-03, -5.82103634e+00, -3.86768103e+00],
        [ 4.81626868e-01,  6.59199810e+00,  1.41322911e+00, ...,
          4.51318836e+00,  1.44330740e+00, -2.11640763e+00],
        [ 5.71519566e+00, -1.73221827e+00, -2.52452445e+00, ...,
         -1.38943791e-01, -4.26732123e-01,  2.83087611e-01],
        ...,
        [ 3.32211196e-01,  4.46825504e+00,  5.59985065e+00, ...,
          1.47309566e+00,  1.84338796e+00,  6.29830837e-01],
        [ 7.48569727e-01, -2.77618349e-01, -9.75789905e-01, ...,
          8.58338699e-02,  7.84621298e-01, -7.33662486e-01],
        [ 3.88537025e+00, -3.19749385e-01, -2.56510758e+00, ...,
         -1.14698482e+00,  1.97455716e+00,  2.41878939e+00]],

       [[-1.22459781e+00,  9.25203860e-01, -3.56069207e+00, ...,
         -6.88289523e-01,  2.72072101e+00,  2.42560315e+00],
        [-2.12065816e+00, -1.31083059e+00,  6.19575642e-02, ...,
         -1.00482607e+00,  6.57156825e-01,  3.69390178e+00],
        [-2.89822197e+00,  3.91538739e-01,  3.01851583e+00, ...,
          3.21430612e+00, -2.96551800e+00,  1.68057680e+00],
        ...,
        [-8.19563568e-01, -1.09326208e+00,  8.15090165e-02, ...,
          9.23141614e-02,  8.50044489e-01, -4.49209958e-02],
        [-1.07429862e+00, -8.72046173e-01, -1.87510759e-01, ...,
         -4.51299906e-01, -3.14396888e-01, -2.45221078e-01],
        [-2.53562868e-01, -6.59026027e-01,  5.62391356e-02, ...,
          2.85453975e-01, -6.72966421e-01, -7.80801594e-01]],

       ...,

       [[ 4.09202874e-01, -2.16477704e+00, -1.33685732e+00, ...,
          6.26661956e-01, -1.69648635e+00,  1.14765966e+00],
        [-6.11324608e-01, -1.46263731e+00,  5.02329171e-01, ...,
          1.89440024e+00,  3.77088010e-01, -4.81952488e-01],
        [-8.16587627e-01, -8.43480527e-01, -6.80629313e-01, ...,
         -3.80911082e-01,  7.75632143e-01,  1.43016458e+00],
        ...,
        [-8.62850428e-01, -5.97268455e-02, -2.32076621e+00, ...,
         -3.94508451e-01, -1.40238211e-01,  8.14994454e-01],
        [-1.88516319e+00, -1.77825320e+00,  2.94450879e+00, ...,
         -2.05334887e-01,  1.53180182e-01,  1.34431088e+00],
        [-1.46972930e+00, -1.01108992e+00,  2.57260472e-01, ...,
         -1.44747949e+00, -1.15590203e+00,  1.05527902e+00]],

       [[ 4.03997850e+00, -7.34005630e-01,  2.56313145e-01, ...,
         -1.19007275e-01, -6.37884915e-01, -4.53511834e-01],
        [ 2.11708620e-01, -7.18083620e-01, -5.61081246e-03, ...,
         -1.00378044e-01, -1.42111838e+00, -9.87072945e-01],
        [ 2.12071109e+00,  1.58937305e-01,  1.37136483e+00, ...,
         -7.05630779e-01, -8.95331025e-01, -6.12974428e-02],
        ...,
        [ 1.64694417e+00,  1.12064087e+00,  2.88810909e-01, ...,
          1.93721259e+00, -5.46663463e-01, -2.24808264e+00],
        [-1.16505757e-01, -1.03227091e+00, -3.96065295e-01, ...,
          9.43170130e-01, -4.81237531e-01, -1.76714098e+00],
        [ 6.67893946e-01, -6.65897429e-01, -8.48173082e-01, ...,
          1.70052081e-01, -1.04281044e+00, -4.57739019e+00]],

       [[-6.04460907e+00,  6.05607557e+00, -2.38166380e+00, ...,
          2.98707813e-01,  3.72313666e+00,  1.61746967e+00],
        [-3.91551423e+00,  7.60167933e+00, -2.10555935e+00, ...,
          3.19988281e-01,  1.98965371e+00,  3.04612070e-01],
        [-4.41497594e-01,  4.27943611e+00, -2.46940994e+00, ...,
         -3.23498416e+00,  1.48617256e+00,  5.61233521e-01],
        ...,
        [ 9.69780743e-01,  1.23910487e+00, -2.21631631e-01, ...,
         -3.57993282e-02, -3.95015627e-01, -3.74649137e-01],
        [ 6.81391358e-01,  2.73380876e-01, -4.88212379e-03, ...,
          5.01719356e-01, -1.70937300e-01,  5.12737215e-01],
        [ 8.33239794e-01, -2.53116101e-01, -8.10810387e-01, ...,
          3.98399562e-01, -3.44598472e-01,  1.71689844e+00]]],
      dtype=float32)

w.shape

(500, 512, 512)

w[[1]].shape

(1, 512, 512)

stg_ps1.shape

(1, 512, 512)

pd.DataFrame(prnu.aligned_cc(stg_ps05,stg_ps1)['ncc'])

	0
0	0.059222

stg_ps1_stg2_ps1_corr=pd.DataFrame()
stg_ps1_stg2_ps1_corr=pd.DataFrame()
stg_ps05_stg_ps1_corr=pd.DataFrame()
stg_ps05_stg2_ps05_corr=pd.DataFrame()
stg2_ps05_stg2_ps1_corr=pd.DataFrame()
stg_ps05_stg2_ps1_corr=pd.DataFrame()

stg_ps05_stg_ps05_corr=pd.DataFrame()
stg_ps1_stg_ps1_corr=pd.DataFrame()
stg2_ps05_stg2_ps05_corr=pd.DataFrame()
stg2_ps1_stg2_ps1_corr=pd.DataFrame()

stg_ps05_normal_corr=pd.DataFrame()
stg_ps1_normal_corr=pd.DataFrame()
stg2_ps05_normal_corr=pd.DataFrame()
stg2_ps1_normal_corr=pd.DataFrame()
normal_normal_corr=pd.DataFrame()

stg_ps05_stg_ps05_corr=pd.DataFrame(prnu.aligned_cc(stg_ps05,stylegan_psi05_w[[0]])['ncc'])
stg_ps05_stg_ps05_corr.columns=["Corr Value"]
for i in range(500):    
    b=pd.DataFrame(prnu.aligned_cc(stg_ps05, stylegan_psi05_w[[i]])['ncc'])
    stg_ps05_stg_ps05_corr=stg_ps05_stg_ps05_corr.append(b)

stg_ps05_stg_ps05_corr=stg_ps05_stg_ps05_corr.iloc[1:]
stg_ps05_stg_ps05_corr.columns=["Corr_Value","Sıl"]
stg_ps05_stg_ps05_corr.drop("Sıl",inplace=True,axis=1)
stg_ps05_stg_ps05_corr.Corr_Value.hist()
print("Correlation mean of StyleGan for psi=0.5 - StyleGan psi=0.5 Pic", stg_ps05_stg_ps05_corr.Corr_Value.mean())

Correlation mean of StyleGan for psi=0.5 - StyleGan psi=0.5 Pic 0.04197201

stg_ps1_stg_ps1_corr=pd.DataFrame(prnu.aligned_cc(stg_ps1,stylegan_psi1_w[[0]])['ncc'])
stg_ps1_stg_ps1_corr.columns=["Corr Value"]
for i in range(500):    
    b=pd.DataFrame(prnu.aligned_cc(stg_ps1, stylegan_psi1_w[[i]])['ncc'])
    stg_ps1_stg_ps1_corr=stg_ps1_stg_ps1_corr.append(b)

stg_ps1_stg_ps1_corr=stg_ps1_stg_ps1_corr.iloc[1:]
stg_ps1_stg_ps1_corr.columns=["Corr_Value","Sıl"]
stg_ps1_stg_ps1_corr.drop("Sıl",inplace=True,axis=1)
stg_ps1_stg_ps1_corr.Corr_Value.hist()
print("Correlation mean of StyleGan for psi=1 - StyleGan psi=1 Pic", stg_ps1_stg_ps1_corr.Corr_Value.mean())

Correlation mean of StyleGan for psi=1 - StyleGan psi=1 Pic 0.039275397

stg2_ps05_stg2_ps05_corr=pd.DataFrame(prnu.aligned_cc(stg2_ps05,stylegan2_psi05_w[[0]])['ncc'])
stg2_ps05_stg2_ps05_corr.columns=["Corr Value"]
for i in range(500):    
    b=pd.DataFrame(prnu.aligned_cc(stg2_ps05, stylegan2_psi05_w[[i]])['ncc'])
    stg2_ps05_stg2_ps05_corr=stg2_ps05_stg2_ps05_corr.append(b)

stg2_ps05_stg2_ps05_corr=stg2_ps05_stg2_ps05_corr.iloc[1:]
stg2_ps05_stg2_ps05_corr.columns=["Corr_Value","Sıl"]
stg2_ps05_stg2_ps05_corr.drop("Sıl",inplace=True,axis=1)
stg2_ps05_stg2_ps05_corr.Corr_Value.hist()
print("Correlation mean of StyleGan2 for psi=0.5 - StyleGan2 psi=0.5 Pic", stg2_ps05_stg2_ps05_corr.Corr_Value.mean())

Correlation mean of StyleGan2 for psi=0.5 - StyleGan2 psi=0.5 Pic 0.04182502

stg2_ps1_stg2_ps1_corr=pd.DataFrame(prnu.aligned_cc(stg2_ps1,stylegan2_psi1_w[[0]])['ncc'])
stg2_ps1_stg2_ps1_corr.columns=["Corr Value"]
for i in range(500):    
    b=pd.DataFrame(prnu.aligned_cc(stg2_ps1, stylegan2_psi1_w[[i]])['ncc'])
    stg2_ps1_stg2_ps1_corr=stg2_ps1_stg2_ps1_corr.append(b)

stg2_ps1_stg2_ps1_corr=stg2_ps1_stg2_ps1_corr.iloc[1:]
stg2_ps1_stg2_ps1_corr.columns=["Corr_Value","Sıl"]
stg2_ps1_stg2_ps1_corr.drop("Sıl",inplace=True,axis=1)
stg2_ps1_stg2_ps1_corr.Corr_Value.hist()
print("Correlation mean of StyleGan2 for psi=1 - StyleGan2 psi=1 Pic", stg2_ps1_stg2_ps1_corr.Corr_Value.mean())

Correlation mean of StyleGan2 for psi=1 - StyleGan2 psi=1 Pic 0.03852026

stg_ps1_stg2_ps05_corr=pd.DataFrame(prnu.aligned_cc(stg_ps1,stylegan2_psi05_w[[0]])['ncc'])
stg_ps1_stg2_ps05_corr.columns=["Corr Value"]
for i in range(500):    
    b=pd.DataFrame(prnu.aligned_cc(stg_ps1, stylegan2_psi05_w[[i]])['ncc'])
    stg_ps1_stg2_ps05_corr=stg_ps1_stg2_ps05_corr.append(b)

stg_ps1_stg2_ps05_corr=stg_ps1_stg2_ps05_corr.iloc[1:]
stg_ps1_stg2_ps05_corr.columns=["Corr_Value","Sıl"]
stg_ps1_stg2_ps05_corr.drop("Sıl",inplace=True,axis=1)
stg_ps1_stg2_ps05_corr.Corr_Value.hist()
print("Correlation mean of StyleGan for psi=1 - StyleGan2 psi=0.5 Pic", stg_ps1_stg2_ps05_corr.Corr_Value.mean())

Correlation mean of StyleGan for psi=1 - StyleGan2 psi=0.5 Pic 0.0013079743

stg_ps05_stg_ps1_corr=pd.DataFrame(prnu.aligned_cc(stg_ps05,stylegan_psi1_w[[0]])['ncc'])
stg_ps05_stg_ps1_corr.columns=["Corr Value"]
for i in range(500):    
    b=pd.DataFrame(prnu.aligned_cc(stg_ps05, stylegan_psi1_w[[i]])['ncc'])
    stg_ps05_stg_ps1_corr=stg_ps05_stg_ps1_corr.append(b)

stg_ps05_stg_ps1_corr=stg_ps05_stg_ps1_corr.iloc[1:]
stg_ps05_stg_ps1_corr.columns=["Corr_Value","Sıl"]
stg_ps05_stg_ps1_corr.drop("Sıl",inplace=True,axis=1)
stg_ps05_stg_ps1_corr.Corr_Value.hist()
print("Correlation mean of StyleGan for psi=0.5 - StyleGan psi=1 Pic", stg_ps05_stg_ps1_corr.Corr_Value.mean())

Correlation mean of StyleGan for psi=0.5 - StyleGan psi=1 Pic 0.0024494517

stg_ps05_stg2_ps1_corr=pd.DataFrame(prnu.aligned_cc(stg_ps05,stylegan2_psi1_w[[0]])['ncc'])
stg_ps05_stg_ps1_corr.columns=["Corr Value"]
for i in range(500):    
    b=pd.DataFrame(prnu.aligned_cc(stg_ps05, stylegan2_psi1_w[[i]])['ncc'])
    stg_ps05_stg2_ps1_corr=stg_ps05_stg2_ps1_corr.append(b)

stg_ps05_stg2_ps1_corr=stg_ps05_stg2_ps1_corr.iloc[1:]

stg_ps05_stg2_ps1_corr.columns=["Corr_Value"]
stg_ps05_stg2_ps1_corr.Corr_Value.hist()
print("Correlation mean of StyleGan for psi=0.5 - StyleGan2 psi=1 Pic", stg_ps05_stg2_ps1_corr.Corr_Value.mean())

Correlation mean of StyleGan for psi=0.5 - StyleGan2 psi=1 Pic 0.0008610556

stg2_ps05_stg2_ps1_corr=pd.DataFrame(prnu.aligned_cc(stg2_ps05,stylegan2_psi1_w[[0]])['ncc'])
stg2_ps05_stg2_ps1_corr.columns=["Corr Value"]
for i in range(500):    
    b=pd.DataFrame(prnu.aligned_cc(stg2_ps05, stylegan2_psi1_w[[i]])['ncc'])
    stg2_ps05_stg2_ps1_corr=stg2_ps05_stg2_ps1_corr.append(b)

stg2_ps05_stg2_ps1_corr=stg2_ps05_stg2_ps1_corr.iloc[1:]
stg2_ps05_stg2_ps1_corr.columns=["Corr_Value","Sıl"]
stg2_ps05_stg2_ps1_corr.drop("Sıl",inplace=True,axis=1)
stg2_ps05_stg2_ps1_corr.columns=["Corr_Value"]
stg2_ps05_stg2_ps1_corr.Corr_Value.hist()
print("Correlation mean of StyleGan2 for psi=0.5 - StyleGan2 psi=1 Pic", stg2_ps05_stg2_ps1_corr.Corr_Value.mean())

Correlation mean of StyleGan2 for psi=0.5 - StyleGan2 psi=1 Pic 0.010030724

stg_ps05_stg2_ps05_corr

stg_ps05_stg2_ps05_corr=pd.DataFrame(prnu.aligned_cc(stg_ps05,stylegan2_psi05_w[[0]])['ncc'])
stg_ps05_stg2_ps05_corr.columns=["Corr Value"]
for i in range(500):    
    b=pd.DataFrame(prnu.aligned_cc(stg_ps05, stylegan2_psi05_w[[i]])['ncc'])
    stg_ps05_stg2_ps05_corr=stg_ps05_stg2_ps05_corr.append(b)

stg_ps05_stg2_ps05_corr=stg_ps05_stg2_ps05_corr.iloc[1:]
stg_ps05_stg2_ps05_corr.columns=["Corr_Value","Sıl"]
stg_ps05_stg2_ps05_corr.drop("Sıl",inplace=True,axis=1)
stg_ps05_stg2_ps05_corr.columns=["Corr_Value"]
stg_ps05_stg2_ps05_corr.Corr_Value.hist()
print("Correlation mean of StyleGan for psi=0.5 - StyleGan psi=0.5 Pic", stg_ps05_stg2_ps05_corr.Corr_Value.mean())

Correlation mean of StyleGan for psi=0.5 - StyleGan psi=0.5 Pic 0.0022130124

stg_ps1_stg2_ps1_corr=pd.DataFrame(prnu.aligned_cc(stg_ps1,stylegan2_psi1_w[[0]])['ncc'])
stg_ps1_stg2_ps1_corr.columns=["Corr Value"]
for i in range(500):    
    b=pd.DataFrame(prnu.aligned_cc(stg_ps1, stylegan2_psi1_w[[i]])['ncc'])
    stg_ps1_stg2_ps1_corr=stg_ps1_stg2_ps1_corr.append(b)

stg_ps1_stg2_ps1_corr=stg_ps1_stg2_ps1_corr.iloc[1:]
stg_ps1_stg2_ps1_corr.columns=["Corr_Value","Sıl"]
stg_ps1_stg2_ps1_corr.drop("Sıl",inplace=True,axis=1)
stg_ps1_stg2_ps1_corr.columns=["Corr_Value"]
stg_ps1_stg2_ps1_corr.Corr_Value.hist()
print("Correlation mean of StyleGan for psi=1 - StyleGan2 psi=1 Pic", stg_ps1_stg2_ps1_corr.Corr_Value.mean())

Correlation mean of StyleGan for psi=1 - StyleGan2 psi=1 Pic 0.00071222236

stg_ps05_normal_corr=pd.DataFrame(prnu.aligned_cc(stg_ps05, w[[0]])['ncc'])
stg_ps05_normal_corr.columns=["Corr Value"]
for i in range(500):    
    b=pd.DataFrame(prnu.aligned_cc(stg_ps05, w[[i]])['ncc'])
    stg_ps05_normal_corr=stg_ps05_normal_corr.append(b)

stg_ps05_normal_corr=stg_ps05_normal_corr.iloc[1:]
stg_ps05_normal_corr.columns=["Corr_Value","Sıl"]
stg_ps05_normal_corr.drop("Sıl",inplace=True,axis=1)
stg_ps05_normal_corr.Corr_Value.hist()
print("Correlation mean of StyleGan for psi=0.5 - Normal Pic", stg_ps05_normal_corr.Corr_Value.mean())

Correlation mean of StyleGan for psi=0.5 - Normal Pic 4.2724823e-05

stg_ps1_normal_corr=pd.DataFrame(prnu.aligned_cc(stg_ps1, w[[0]])['ncc'])
stg_ps1_normal_corr.columns=["Corr Value"]
for i in range(500):    
    b=pd.DataFrame(prnu.aligned_cc(stg_ps1, w[[i]])['ncc'])
    stg_ps1_normal_corr=stg_ps1_normal_corr.append(b)

stg_ps1_normal_corr=stg_ps1_normal_corr.iloc[1:]
stg_ps1_normal_corr.columns=["Corr_Value","Sıl"]
stg_ps1_normal_corr.drop("Sıl",inplace=True,axis=1)
stg_ps1_normal_corr.Corr_Value.hist()
print("Correlation mean of StyleGan for psi=1 - Normal Pic", stg_ps1_normal_corr.Corr_Value.mean())

Correlation mean of StyleGan for psi=1 - Normal Pic 0.00028177295

print("- This shows the correlation between 500 normal images and 500 GAN images' average fingerprint.")
stg_ps1_normal_corr

- This shows the correlation between 500 normal images and 500 GAN images' average fingerprint.

	Corr_Value
0	-0.001067
0	-0.003791
0	0.002194
0	0.001792
0	0.002685
...	...
0	0.001814
0	-0.003174
0	-0.001017
0	0.001692
0	-0.003464

500 rows × 1 columns

stg2_ps05_normal_corr=pd.DataFrame(prnu.aligned_cc(stg2_ps05, w[[0]])['ncc'])
stg2_ps05_normal_corr.columns=["Corr Value"]
for i in range(500):    
    b=pd.DataFrame(prnu.aligned_cc(stg2_ps05, w[[i]])['ncc'])
    stg2_ps05_normal_corr=stg2_ps05_normal_corr.append(b)

stg2_ps05_normal_corr=stg2_ps05_normal_corr.iloc[1:]
stg2_ps05_normal_corr.columns=["Corr_Value","Sıl"]
stg2_ps05_normal_corr.drop("Sıl",inplace=True,axis=1)
stg2_ps05_normal_corr.Corr_Value.hist()
print("Correlation mean of StyleGan2 for psi=0.5 - Normal Pic", stg2_ps05_normal_corr.Corr_Value.mean())

Correlation mean of StyleGan2 for psi=0.5 - Normal Pic -1.344089e-05

stg2_ps05_normal_corr

	Corr_Value
0	-0.002844
0	-0.001596
0	-0.000229
0	0.001705
0	0.001134
...	...
0	-0.001822
0	-0.002067
0	0.001924
0	0.000356
0	0.001591

500 rows × 1 columns

stg2_ps1_normal_corr=pd.DataFrame(prnu.aligned_cc(stg2_ps1, w[[0]])['ncc'])
stg2_ps1_normal_corr.columns=["Corr Value"]
for i in range(500):    
    b=pd.DataFrame(prnu.aligned_cc(stg2_ps1, w[[i]])['ncc'])
    stg2_ps1_normal_corr=stg2_ps1_normal_corr.append(b)

stg2_ps1_normal_corr=stg2_ps1_normal_corr.iloc[1:]
stg2_ps1_normal_corr.columns=["Corr_Value","Sıl"]
stg2_ps1_normal_corr.drop("Sıl",inplace=True,axis=1)
stg2_ps1_normal_corr.Corr_Value.hist()
print("Correlation mean of StyleGan2 for psi=1 - Normal Pic", stg2_ps1_normal_corr.Corr_Value.mean())

Correlation mean of StyleGan2 for psi=1 - Normal Pic -7.05533e-05

stg2_ps1_normal_corr

	Corr_Value
0	-0.006473
0	0.003326
0	0.000535
0	-0.002722
0	0.001405
...	...
0	0.000236
0	-0.003573
0	0.002968
0	-0.002358
0	0.003772

500 rows × 1 columns

normal_normal_corr=pd.DataFrame(prnu.aligned_cc(c, w[[0]])['ncc'])
normal_normal_corr.columns=["Corr Value"]
for i in range(500):    
    b=pd.DataFrame(prnu.aligned_cc(c, w[[i]])['ncc'])
    normal_normal_corr=normal_normal_corr.append(b)
    
normal_normal_corr=normal_normal_corr.iloc[1:]
normal_normal_corr.columns=["Corr_Value","Sıl"]
normal_normal_corr.drop("Sıl",inplace=True,axis=1)
normal_normal_corr.Corr_Value.hist()
print("Correlation mean of Normal-Normal Pic", normal_normal_corr.Corr_Value.mean())

Correlation mean of Normal-Normal Pic 0.03800994

print("Correlation mean of StyleGan for psi=0.5 - Normal Pic", stg_ps05_normal_corr.Corr_Value.mean())
print("Correlation mean of StyleGan for psi=1 - Normal Pic", stg_ps1_normal_corr.Corr_Value.mean())
print("Correlation mean of StyleGan2 for psi=0.5 - Normal Pic", stg2_ps05_normal_corr.Corr_Value.mean())
print("Correlation mean of StyleGan2 for psi=1 - Normal Pic", stg2_ps1_normal_corr.Corr_Value.mean())
print("Correlation mean of Normal-Normal Pic", normal_normal_corr.Corr_Value.mean())

Correlation mean of StyleGan for psi=0.5 - Normal Pic 4.2724823e-05
Correlation mean of StyleGan for psi=1 - Normal Pic 0.00028177295
Correlation mean of StyleGan2 for psi=0.5 - Normal Pic -1.344089e-05
Correlation mean of StyleGan2 for psi=1 - Normal Pic -7.05533e-05
Correlation mean of Normal-Normal Pic 0.03800994

Results

• As we expected, we get higher correlation for StyleGan as we increase psi from 0.5 to 1.0.
• However, we get less correlation for StyleGan2 than StyleGan when we increase psi from 0.5 to 1.0.
• Finally, normal - normal correlation much higher than both Gans.

stg_ps05_stg2_ps05_corr

	Corr_Value
0	0.001487
0	-0.001323
0	0.005130
0	0.010521
0	-0.005079
...	...
0	-0.002846
0	0.007465
0	0.005167
0	0.001961
0	-0.001741

500 rows × 1 columns

import seaborn as sns
sns.distplot(stg_ps1_normal_corr["Corr_Value"] , color="blue", label="Normal & StyleGAN psi=1",bins=300)
sns.distplot(stg2_ps1_normal_corr["Corr_Value"] , color="red", label="Normal & StyleGAN2 psi=1",bins=300)
sns.distplot(normal_normal_corr["Corr_Value"] , color="green", label="Normal & Normal",bins=300)

sns.distplot(stg_ps05_stg2_ps05_corr["Corr_Value"] , color="orange", label="Stg_ps05 & Stg2_ps05 ",bins=300)
sns.distplot(stg_ps05_stg2_ps1_corr["Corr_Value"] , color="red", label="Stg_ps05 & Stg2_ps1 ",bins=300)
sns.distplot(stg_ps1_stg2_ps1_corr["Corr_Value"] , color="cyan", label="Stg_ps1 & Stg2_ps1 ",bins=300)
sns.distplot(stg_ps1_stg2_ps05_corr["Corr_Value"] , color="plum", label="Stg_ps1 & Stg2_ps05 ",bins=300)
sns.distplot(stg2_ps05_stg2_ps1_corr["Corr_Value"] , color="black", label="Stg2_ps1 & Stg2_ps05 ",bins=300)
sns.distplot(stg2_ps1_stg2_ps1_corr["Corr_Value"] , color="brown", label="Stg_ps05 & Stg2_ps1 ",bins=300)
sns.distplot(stg2_ps05_stg2_ps05_corr["Corr_Value"] , color="indigo", label="Stg2_ps05 & Stg2_ps05 ",bins=300)
sns.distplot(stg_ps05_stg_ps05_corr["Corr_Value"] , color="sienna", label="Stg_ps05 & Stg_ps05 ",bins=300)
sns.distplot(stg_ps1_stg_ps1_corr["Corr_Value"] , color="palegreen", label="Stg_ps1 & Stg_ps1 ",bins=300)
plt.legend()

plt.title('Histogram of correlations')

C:\Users\user\anaconda3\envs\tf15\lib\site-packages\seaborn\distributions.py:2551: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).
  warnings.warn(msg, FutureWarning)
C:\Users\user\anaconda3\envs\tf15\lib\site-packages\seaborn\distributions.py:2551: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).
  warnings.warn(msg, FutureWarning)
C:\Users\user\anaconda3\envs\tf15\lib\site-packages\seaborn\distributions.py:2551: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).
  warnings.warn(msg, FutureWarning)
C:\Users\user\anaconda3\envs\tf15\lib\site-packages\seaborn\distributions.py:2551: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).
  warnings.warn(msg, FutureWarning)
C:\Users\user\anaconda3\envs\tf15\lib\site-packages\seaborn\distributions.py:2551: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).
  warnings.warn(msg, FutureWarning)
C:\Users\user\anaconda3\envs\tf15\lib\site-packages\seaborn\distributions.py:2551: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).
  warnings.warn(msg, FutureWarning)
C:\Users\user\anaconda3\envs\tf15\lib\site-packages\seaborn\distributions.py:2551: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).
  warnings.warn(msg, FutureWarning)
C:\Users\user\anaconda3\envs\tf15\lib\site-packages\seaborn\distributions.py:2551: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).
  warnings.warn(msg, FutureWarning)
C:\Users\user\anaconda3\envs\tf15\lib\site-packages\seaborn\distributions.py:2551: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).
  warnings.warn(msg, FutureWarning)
C:\Users\user\anaconda3\envs\tf15\lib\site-packages\seaborn\distributions.py:2551: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).
  warnings.warn(msg, FutureWarning)
C:\Users\user\anaconda3\envs\tf15\lib\site-packages\seaborn\distributions.py:2551: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).
  warnings.warn(msg, FutureWarning)
C:\Users\user\anaconda3\envs\tf15\lib\site-packages\seaborn\distributions.py:2551: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).
  warnings.warn(msg, FutureWarning)

Text(0.5, 1.0, 'Histogram of correlations')

As seen from histograms, there are higher correlation between same classes.

Creating the heat map of correlations.

stg_ps05_normal_max = stg_ps05_normal_corr.mean()
stg_ps1_normal_max = stg_ps1_normal_corr.mean()
stg2_ps05_normal_max = stg2_ps05_normal_corr.mean()
stg2_ps1_normal_max = stg2_ps1_normal_corr.mean()
normal_normal_max = normal_normal_corr.mean()
stg_ps1_stg2_ps1_max = stg_ps1_stg2_ps1_corr.mean()
stg_ps1_stg2_ps05_max = stg_ps1_stg2_ps05_corr.mean()
stg_ps05_stg_ps1_max = stg_ps05_stg_ps1_corr.mean()
stg_ps05_stg2_ps05_max = stg_ps05_stg2_ps05_corr.mean()
stg2_ps05_stg2_ps1_max = stg2_ps05_stg2_ps1_corr.mean()
stg_ps05_stg2_ps1_max = stg_ps05_stg2_ps1_corr.mean()

stg_ps05_stg_ps05_max = stg_ps05_stg_ps05_corr.mean()
stg_ps1_stg_ps1_max = stg_ps1_stg_ps1_corr.mean()
stg2_ps05_stg2_ps05_max = stg2_ps05_stg2_ps05_corr.mean()
stg2_ps1_stg2_ps1_max = stg2_ps1_stg2_ps1_corr.mean()

stg_ps05_dt=[stg_ps05_normal_max[0],stg_ps05_stg_ps1_max[0],stg_ps05_stg2_ps05_max[0],stg_ps05_stg2_ps1_max[0]]

stg_ps05_dt=[stg_ps05_stg_ps05_max[0],stg_ps05_normal_max[0],stg_ps05_stg_ps1_max[0],stg_ps05_stg2_ps05_max[0],stg_ps05_stg2_ps1_max[0]]
from pandas import DataFrame
stg_ps05_dt=DataFrame(stg_ps05_dt).T
stg_ps05_dt.columns=["stg_ps05","Normal","stg_ps1","stg2_ps05","stg2_ps1"]

stg_ps05_dt

	stg_ps05	Normal	stg_ps1	stg2_ps05	stg2_ps1
0	0.041972	0.000043	0.002449	0.002213	0.000861

stg_ps1_dt=[stg_ps05_stg_ps1_max[0],stg_ps1_normal_max[0],stg_ps1_stg_ps1_max[0],stg_ps1_stg2_ps05_max[0],stg_ps1_stg2_ps1_max[0]]
from pandas import DataFrame
stg_ps1_dt=DataFrame(stg_ps1_dt).T
stg_ps1_dt.columns=["stg_ps05","Normal","stg_ps1","stg2_ps05","stg2_ps1"]
stg_ps1_dt

	stg_ps05	Normal	stg_ps1	stg2_ps05	stg2_ps1
0	0.002449	0.000282	0.039275	0.001308	0.000712

stg2_ps05_normal_max[0]

-1.344089e-05

stg2_ps05_dt=[stg_ps05_stg2_ps05_max[0],stg2_ps05_normal_max[0],stg2_ps05_stg2_ps1_max[0],stg2_ps05_stg2_ps05_max[0],stg2_ps05_stg2_ps1_max[0]]
from pandas import DataFrame
stg2_ps05_dt=DataFrame(stg2_ps05_dt).T
stg2_ps05_dt.columns=["stg_ps05","Normal","stg_ps1","stg2_ps05","stg2_ps1"]
stg2_ps05_dt

	stg_ps05	Normal	stg_ps1	stg2_ps05	stg2_ps1
0	0.002213	-0.000013	0.010031	0.041825	0.010031

stg2_ps1_dt=[stg_ps05_stg2_ps1_max[0],stg2_ps1_normal_max[0],stg_ps1_stg2_ps1_max[0],stg2_ps05_stg2_ps1_max[0],stg2_ps1_stg2_ps1_max[0]]
from pandas import DataFrame
stg2_ps1_dt=DataFrame(stg2_ps1_dt).T
stg2_ps1_dt.columns=["stg_ps05","Normal","stg_ps1","stg2_ps05","stg2_ps1"]
stg2_ps1_dt

	stg_ps05	Normal	stg_ps1	stg2_ps05	stg2_ps1
0	0.000861	-0.000071	0.000712	0.010031	0.03852

normal_dt=[stg_ps05_normal_max[0],normal_normal_max[0],stg_ps1_normal_max[0],stg2_ps05_normal_max[0],stg2_ps1_normal_max[0]]
from pandas import DataFrame
normal_dt=DataFrame(normal_dt).T
normal_dt.columns=["stg_ps05","Normal","stg_ps1","stg2_ps05","stg2_ps1"]
normal_dt

	stg_ps05	Normal	stg_ps1	stg2_ps05	stg2_ps1
0	0.000043	0.03801	0.000282	-0.000013	-0.000071

corr_matrix=pd.concat([stg_ps05_dt,normal_dt,stg_ps1_dt,stg2_ps05_dt,stg2_ps1_dt])

ındex=pd.DataFrame(["stg_ps05","normal","stg_ps1","stg2_ps05","stg2_ps1"])
ındex.columns=["Corr"]

ındex.set_index("Corr")

Corr
stg_ps05
normal
stg_ps1
stg2_ps05
stg2_ps1

corr_matrix.set_index(ındex["Corr"])

	stg_ps05	Normal	stg_ps1	stg2_ps05	stg2_ps1
Corr
stg_ps05	0.041972	0.000043	0.002449	0.002213	0.000861
normal	0.000043	0.038010	0.000282	-0.000013	-0.000071
stg_ps1	0.002449	0.000282	0.039275	0.001308	0.000712
stg2_ps05	0.002213	-0.000013	0.010031	0.041825	0.010031
stg2_ps1	0.000861	-0.000071	0.000712	0.010031	0.038520

sns.heatmap(corr_matrix, xticklabels=corr_matrix.columns, yticklabels=corr_matrix.columns,annot=True, center=0,vmin=-1, vmax=1)

<AxesSubplot:>

This correlation matrix shows the same results as the histogram of correlation above.

normal_normal_corr

	Corr_Value
0	0.041360
0	0.033125
0	0.033120
0	0.035139
0	0.061904
...	...
0	0.039467
0	0.044159
0	0.032853
0	0.059877
0	0.037271

500 rows × 1 columns

# Import library and dataset
import seaborn as sns

sns.set(rc={'figure.figsize':(11.7,8.27)})


# Method 1: on the same Axis
sns.distplot(stg_ps05_normal_corr["Corr_Value"] , color="blue", label="StyleGAN psi=0.5",bins=300)
sns.distplot(stg2_ps05_normal_corr["Corr_Value"] , color="red", label="StyleGAN2 psi=0.5",bins=300)
sns.distplot(stg_ps1_normal_corr["Corr_Value"] , color="black", label="StyleGAN psi=1",bins=300)
sns.distplot(stg2_ps1_normal_corr["Corr_Value"] , color="orange", label="StyleGAN2 psi=1",bins=300)
sns.distplot(normal_normal_corr["Corr_Value"] , color="green", label="Normal", bins=300)
plt.legend()

plt.title('Histogram of correlation between Normal Pictures')
plt.xlabel('Corr_Values')


#sns.plt.show()

C:\Users\user\anaconda3\envs\tf15\lib\site-packages\seaborn\distributions.py:2551: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).
  warnings.warn(msg, FutureWarning)
C:\Users\user\anaconda3\envs\tf15\lib\site-packages\seaborn\distributions.py:2551: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).
  warnings.warn(msg, FutureWarning)
C:\Users\user\anaconda3\envs\tf15\lib\site-packages\seaborn\distributions.py:2551: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).
  warnings.warn(msg, FutureWarning)
C:\Users\user\anaconda3\envs\tf15\lib\site-packages\seaborn\distributions.py:2551: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).
  warnings.warn(msg, FutureWarning)
C:\Users\user\anaconda3\envs\tf15\lib\site-packages\seaborn\distributions.py:2551: FutureWarning: `distplot` is a deprecated function and will be removed in a future version. Please adapt your code to use either `displot` (a figure-level function with similar flexibility) or `histplot` (an axes-level function for histograms).
  warnings.warn(msg, FutureWarning)

Text(0.5, 0, 'Corr_Values')

We defined that the best threshold value should be the intersection point of GAN and Normal pictures.

Classification

normal_normal_corr["Label"]=1

stg_ps05_normal_corr.reset_index(inplace=True)
stg_ps05_normal_corr.drop("index",inplace=True,axis=1)

stg2_ps05_normal_corr.reset_index(inplace=True)
stg2_ps05_normal_corr.drop("index",inplace=True,axis=1)

stg2_ps1_normal_corr.reset_index(inplace=True)
stg2_ps1_normal_corr.drop("index",inplace=True,axis=1)

stg_ps1_normal_corr.reset_index(inplace=True)
stg_ps1_normal_corr.drop("index",inplace=True,axis=1)

stg1_2=pd.concat([stg_ps05_normal_corr,stg2_ps05_normal_corr,stg2_ps1_normal_corr,stg_ps1_normal_corr],axis=1)

stg_ps05_normal_corr

	Corr_Value
0	-0.000878
1	-0.000172
2	-0.002831
3	-0.000212
4	-0.003578
...	...
495	0.001973
496	-0.002398
497	0.002643
498	0.001534
499	-0.000134

500 rows × 1 columns

stg1_2=pd.concat([stg_ps05_normal_corr,stg2_ps05_normal_corr,stg2_ps1_normal_corr,stg_ps1_normal_corr],axis=1)

stg1_2

	Corr_Value	Corr_Value	Corr_Value	Corr_Value
0	-0.000878	-0.002844	-0.006473	-0.001067
1	-0.000172	-0.001596	0.003326	-0.003791
2	-0.002831	-0.000229	0.000535	0.002194
3	-0.000212	0.001705	-0.002722	0.001792
4	-0.003578	0.001134	0.001405	0.002685
...	...	...	...	...
495	0.001973	-0.001822	0.000236	0.001814
496	-0.002398	-0.002067	-0.003573	-0.003174
497	0.002643	0.001924	0.002968	-0.001017
498	0.001534	0.000356	-0.002358	0.001692
499	-0.000134	0.001591	0.003772	-0.003464

500 rows × 4 columns

stg_mean=pd.DataFrame(stg1_2["Corr_Value"].T.mean())
stg_mean.columns=["Corr_mean"]
stg_mean

	Corr_mean
0	-0.002815
1	-0.000558
2	-0.000083
3	0.000141
4	0.000411
...	...
495	0.000550
496	-0.002803
497	0.001629
498	0.000306
499	0.000441

500 rows × 1 columns

normal_normal_corr["Label"]=1
stg_mean["Label"]=0
normal_normal_corr["Prediction"]=1
normal_normal_corr.columns=["Corr_mean","Label","Prediction"]

df=pd.concat([stg_mean,normal_normal_corr],axis=0)

print(df.head())
print(df.shape)

   Corr_mean  Label  Prediction
0  -0.002815      0         NaN
1  -0.000558      0         NaN
2  -0.000083      0         NaN
3   0.000141      0         NaN
4   0.000411      0         NaN
(1000, 3)

From 500 GAN and 500 Normal pictures we labeled them as 0 and 1 accordingly. Prediction column defined as randomly.

df["Prediction"]=0
def state_process(cell_value):
    if cell_value >0:     #Threshold is 0.
        return 1
    else:
        return 0    
df.Prediction = df.Corr_mean.apply(state_process)

df

	Corr_mean	Label	Prediction
0	-0.002815	0	0
1	-0.000558	0	0
2	-0.000083	0	0
3	0.000141	0	1
4	0.000411	0	1
...	...	...	...
0	0.039467	1	1
0	0.044159	1	1
0	0.032853	1	1
0	0.059877	1	1
0	0.037271	1	1

1000 rows × 3 columns

df.Prediction.value_counts()

1    768
0    232
Name: Prediction, dtype: int64

from sklearn.model_selection import cross_val_score
from sklearn import metrics

confusion_matrix=metrics.confusion_matrix(df.Label,df.Prediction)
confusion_matrix

array([[232, 268],
       [  0, 500]], dtype=int64)

auc_roc=metrics.roc_auc_score(df.Label,df.Prediction)
auc_roc

0.732

If we set threshold as 0, we find AUC as 0.732. However, we predicted 100% correctly of all the normal pictures.

false_positive_rate, true_positive_rate, thresholds = roc_curve(df["Label"], df["Prediction"])
roc_auc = auc(false_positive_rate, true_positive_rate)
roc_auc

import matplotlib.pyplot as plt
plt.figure(figsize=(10,10))
plt.title('Receiver Operating Characteristic')
plt.plot(false_positive_rate,true_positive_rate, color='red',label = 'AUC = %0.2f' % roc_auc)
plt.legend(loc = 'lower right')
plt.plot([0, 1], [0, 1],linestyle='--')
plt.axis('tight')
plt.ylabel('True Positive Rate')
plt.xlabel('False Positive Rate')

Text(0.5, 0, 'False Positive Rate')

Threshold tuning

score=[]
threshold=[]
for i in range(-5,50):
    i=i/1000
    threshold.append(i)
    def state_process(cell_value):
        if cell_value >i:
            return 1
        else:
            return 0    
    df.Prediction = df.Corr_mean.apply(state_process)
    auc_roc=metrics.roc_auc_score(df.Label,df.Prediction)
    score.append(auc_roc)
    

th_score=pd.DataFrame([score,threshold]).T
th_score.columns=["score","Threshold"]

plt.plot(th_score.Threshold,th_score.score, color='red')
plt.legend(loc = 'lower right')
plt.axis('tight')
plt.ylabel('AUC Score')
plt.xlabel('Threshold')

No handles with labels found to put in legend.

Text(0.5, 0, 'Threshold')

plt.margins(x=-0.1, y=0.010)  # Values in (-0.5, 0.0) zooms in to center
plt.plot(th_score.Threshold, (th_score.score))

[<matplotlib.lines.Line2D at 0x1269e19a888>]

As can be seen from the graph above, best threshold value is 0.005.

Test Picture

stylegan2_psi1 = np.array(sorted(glob(r'test\data\stylegan2_psi1.0\*.PNG')))
stylegan2_psi1_device = np.array([os.path.split(i)[1].split('0',1 )[0] for i in stylegan2_psi1])[:600]
sample=stylegan2_psi1[599:600]

sample

array(['test\\data\\stylegan2_psi1.0\\000599.png'], dtype='<U37')

from PIL import Image

basewidth = 512
img_sample = Image.open(sample[0])
wpercent = (basewidth/float(img_sample.size[0]))
hsize = int((float(img_sample.size[1])*float(wpercent)))
img_sample = img_sample.resize((basewidth,hsize), Image.ANTIALIAS)
img_sample

sample[0]

'test\\data\\stylegan2_psi1.0\\000599.png'

We computed fingerprints for test picture.

sample_img=prnu.extract_single(np.array(img_sample)) #extract sıngle pıcture resıduals
sample_img = np.stack(sample_img, 0)
print(sample_img.shape)
from numpy import zeros, newaxis
sample_img = sample_img[newaxis,:, :]
print(sample_img.shape)

(512, 512)
(1, 512, 512)

Firstly, we correlated test picture's fingerprint with 500 normal picture residual. Then we take mean of this normalized cross-correlation and see if it is higher or lower than the threshold.

pred=pd.DataFrame(prnu.aligned_cc(sample_img,c)['ncc']) #ncc between test resıduals and normal pıctures fıngerprınts
print(pred, "--> This value is mean of NCC of 500 normal pictures fingerprints and test picture's residuals.")

          0
0 -0.003205 --> This value is mean of NCC of 500 normal pictures fingerprints and test picture's residuals.

if pred.values >0.001:  #threshold
    print("Prediction is Normal Picture")
else:
    print("Prediction is GAN Picture") 

Prediction is GAN Picture

Lasly, we selected a random StyleGAN2 picture with psi = 1.0 and threshold = 0,001, predicted as GAN picture, which is correct.

Tests

cat = np.array(sorted(glob(r'test\data\stylegan2_psi1.0\000918.PNG')))

basewidth = 512
img_sample_cat = Image.open(cat[0])
wpercent = (basewidth/float(img_sample_cat.size[0]))
hsize = 512
img_sample_cat = img_sample_cat.resize((basewidth,hsize), Image.ANTIALIAS)
img_sample_cat

np.array(img_sample_cat).shape

(512, 512, 3)

sample_res=prnu.extract_single(np.array(img_sample_cat)) #extract sıngle pıcture resıduals
sample_res = np.stack(sample_res, 0)
print(sample_res.shape)
from numpy import zeros, newaxis
sample_res = sample_res[newaxis,:, :]
pred=pd.DataFrame(prnu.aligned_cc(sample_res,c)['ncc']) #ncc between test resıduals and normal pıctures fıngerprınts
print(pred, "--> This value is mean of NCC of 500 normal pictures fingerprints and test picture's residuals.")

(512, 512)
          0
0 -0.004422 --> This value is mean of NCC of 500 normal pictures fingerprints and test picture's residuals.

if pred.values >0.001: #threshold
    print("Prediction is Normal Picture")
else:
    print("Prediction is GAN Picture") 

Prediction is GAN Picture

2) Gan Test Pic

cat = np.array(sorted(glob(r'test\data\stylegan2_psi1.0\000938.PNG')))
basewidth = 512
img_sample_cat = Image.open(cat[0])
wpercent = (basewidth/float(img_sample_cat.size[0]))
hsize = 512
img_sample_cat = img_sample_cat.resize((basewidth,hsize), Image.ANTIALIAS)
img_sample_cat

sample_res=prnu.extract_single(np.array(img_sample_cat)) #extract sıngle pıcture resıduals
sample_res = np.stack(sample_res, 0)
print(sample_res.shape)
from numpy import zeros, newaxis
sample_res = sample_res[newaxis,:, :]
pred=pd.DataFrame(prnu.aligned_cc(sample_res,c)['ncc']) #ncc between test resıduals and normal pıctures fıngerprınts
print(pred, "--> This value is mean of NCC of 500 normal pictures fingerprints and test picture's residuals.")
if pred.values >0.001:  #threshold
    print("Prediction is Normal Picture")
else:
    print("Prediction is GAN Picture") 

(512, 512)
          0
0 -0.002328 --> This value is mean of NCC of 500 normal pictures fingerprints and test picture's residuals.
Prediction is GAN Picture

3) Normal Pic Test

cat = np.array(sorted(glob(r'test\data\archieve2\53122.PNG')))
basewidth = 512
img_sample_cat = Image.open(cat[0])
wpercent = (basewidth/float(img_sample_cat.size[0]))
hsize = 512
img_sample_cat = img_sample_cat.resize((basewidth,hsize), Image.ANTIALIAS)
img_sample_cat

sample_res=prnu.extract_single(np.array(img_sample_cat)) #extract sıngle pıcture resıduals
sample_res = np.stack(sample_res, 0)
print(sample_res.shape)
from numpy import zeros, newaxis
sample_res = sample_res[newaxis,:, :]
pred=pd.DataFrame(prnu.aligned_cc(sample_res,c)['ncc']) #ncc between test resıduals and normal pıctures fıngerprınts
print(pred, "--> This value is mean of NCC of 500 normal pictures fingerprints and test picture's residuals.")
if pred.values >0.005:  #threshold
    print("Prediction is Normal Picture")
else:
    print("Prediction is GAN Picture") 

(512, 512)
          0
0  0.002004 --> This value is mean of NCC of 500 normal pictures fingerprints and test picture's residuals.
Prediction is GAN Picture

Our test pictures

cat = np.array(sorted(glob(r'test\data\mr_ali2.JPG')))
basewidth = 512
img_sample_cat = Image.open(cat[0])
wpercent = (basewidth/float(img_sample_cat.size[0]))
hsize = 512
img_sample_cat = img_sample_cat.resize((basewidth,hsize), Image.ANTIALIAS)
img_sample_cat

sample_res=prnu.extract_single(np.array(img_sample_cat)) #extract sıngle pıcture resıduals
sample_res = np.stack(sample_res, 0)
print(sample_res.shape)
from numpy import zeros, newaxis
sample_res = sample_res[newaxis,:, :]
pred=pd.DataFrame(prnu.aligned_cc(sample_res,c)['ncc']) #ncc between test resıduals and normal pıctures fıngerprınts
print(pred, "--> This value is mean of NCC of 500 normal pictures fingerprints and test picture's residuals.")
if pred.values >0.001:  #threshold
    print("Prediction is Normal Picture")
else:
    print("Prediction is GAN Picture") 

(512, 512)
          0
0  0.002741 --> This value is mean of NCC of 500 normal pictures fingerprints and test picture's residuals.
Prediction is Normal Picture

cat = np.array(sorted(glob(r'test\data\test\normal\05457.PNG')))
basewidth = 512
img_sample_cat = Image.open(cat[0])
wpercent = (basewidth/float(img_sample_cat.size[0]))
hsize = 512
img_sample_cat = img_sample_cat.resize((basewidth,hsize), Image.ANTIALIAS)
img_sample_cat

sample_res=prnu.extract_single(np.array(img_sample_cat)) #extract sıngle pıcture resıduals
sample_res = np.stack(sample_res, 0)
print(sample_res.shape)
from numpy import zeros, newaxis
sample_res = sample_res[newaxis,:, :]
pred=pd.DataFrame(prnu.aligned_cc(sample_res,c)['ncc']) #ncc between test resıduals and normal pıctures fıngerprınts
print(pred, "--> This value is mean of NCC of 500 normal pictures fingerprints and test picture's residuals.")
if pred.values >0.001:  #threshold
    print("Prediction is Normal Picture")
else:
    print("Prediction is GAN Picture") 

(512, 512)
          0
0  0.001506 --> This value is mean of NCC of 500 normal pictures fingerprints and test picture's residuals.
Prediction is Normal Picture

sonucc=pd.DataFrame()
sonucc["Predıctıon"]=0
sonucc

	Predıctıon

if pred.values >0.001:  #threshold
    sonucc["Predıctıon"].loc[0]=0
else:
    sonucc["Predıctıon"].loc[0]=1
    

sonucc["Predıctıon"].loc[1]=1
sonucc

	Predıctıon

pred.append(pred)

	0
0	0.001192
0	0.001192

According to the AUC graph we should've select the threshold as 0.005. However, we get better results when we select threshold as 0.001, so we stick to that value.

Test

nat_dirlist_test = np.array(sorted(glob(r'test\data\test\normal\*.PNG')))[:300]
nat_device_test = np.array([os.path.split(i)[1].split('0', 1)[0] for i in nat_dirlist_test])[:300]

stylegan_psi05_test = np.array(sorted(glob(r'test\data\test\stylegan_psi0.5\*.PNG')))[:300]
stylegan_psi05_device_test = np.array([os.path.split(i)[1].rsplit('_', 1)[0] for i in stylegan_psi05_test])[:300]

stylegan_psi1_test = np.array(sorted(glob(r'test\data\test\stylegan_psi1.0\*.PNG')))[:300]
stylegan_psi1_device_test = np.array([os.path.split(i)[1].rsplit('_', 1)[0] for i in stylegan_psi1_test])[:300]

stylegan2_psi05_test = np.array(sorted(glob(r'test\data\test\stylegan2_psi0.5\*.PNG')))[:300]
stylegan2_psi05_device_test = np.array([os.path.split(i)[1].split('0', 1)[0] for i in stylegan2_psi05_test])[:300]
stylegan2_psi05_test=stylegan2_psi05_test[:300]

stylegan2_psi1_test = np.array(sorted(glob(r'test\data\test\stylegan2_psi1.0\*.PNG')))[:300]
stylegan2_psi1_device_test = np.array([os.path.split(i)[1].split('0',1 )[0] for i in stylegan2_psi1_test])[:300]
stylegan2_psi1_test=stylegan2_psi1_test[:300]

nat_dirlist_test[3]

'test\\data\\test\\normal\\05307.png'

predıctıon_values=[]
corr=[]
for i in range(0,299):    
    img_sample_cat = nat_dirlist_test[i]
    basewidth = 512
    img_sample_cat = Image.open(img_sample_cat)
    wpercent = (basewidth/float(img_sample_cat.size[0]))
    hsize = 512
    img_sample_cat = img_sample_cat.resize((basewidth,hsize), Image.ANTIALIAS)
    
    sample_res=prnu.extract_single(np.array(img_sample_cat)) #extract sıngle pıcture resıduals
    sample_res = np.stack(sample_res, 0)
    from numpy import zeros, newaxis
    sample_res = sample_res[newaxis,:, :]
    pred=pd.DataFrame(prnu.aligned_cc(sample_res,c)['ncc']) #ncc between test resıduals and normal pıctures fıngerprınts
    if pred.values >0.001:  #threshold
        predıctıon_values.append(0)
        corr.append(pred.values)
    else:
        predıctıon_values.append(1)
        corr.append(pred.values)
predıctıon_values=pd.DataFrame(predıctıon_values)
predıctıon_values["label"]=1
predıctıon_values.columns=["Pred","label"]
predıctıon_values.head()
from sklearn.metrics import accuracy_score
print("Normal Accuracy Score:",accuracy_score(predıctıon_values.label,predıctıon_values.Pred))
print("Normal correlation mean: ", np.mean(corr))

Normal Accuracy Score: 0.5986622073578596
Normal correlation mean:  0.00022095379

predıctıon_values=[]
corr=[]
for i in range(0,299):    
    img_sample_cat = stylegan_psi05_test[i]
    basewidth = 512
    img_sample_cat = Image.open(img_sample_cat)
    wpercent = (basewidth/float(img_sample_cat.size[0]))
    hsize = 512
    img_sample_cat = img_sample_cat.resize((basewidth,hsize), Image.ANTIALIAS)
    
    sample_res=prnu.extract_single(np.array(img_sample_cat)) #extract sıngle pıcture resıduals
    sample_res = np.stack(sample_res, 0)
    from numpy import zeros, newaxis
    sample_res = sample_res[newaxis,:, :]
    pred=pd.DataFrame(prnu.aligned_cc(sample_res,c)['ncc']) #ncc between test resıduals and normal pıctures fıngerprınts
    if pred.values >0.001:  #threshold
        predıctıon_values.append(0)
        corr.append(pred.values)
    else:
        predıctıon_values.append(1)
        corr.append(pred.values)
predıctıon_values=pd.DataFrame(predıctıon_values)
predıctıon_values["label"]=1
predıctıon_values.columns=["Pred","label"]
predıctıon_values.head()
from sklearn.metrics import accuracy_score
print("Style GAN psi05 Accuracy Score:",accuracy_score(predıctıon_values.label,predıctıon_values.Pred))
print("Style GAN psi05 correlation mean: ", np.mean(corr))

Style GAN psi05 Accuracy Score: 0.5986622073578596
Style GAN psi05 correlation mean:  0.00030151452

predıctıon_values=[]
corr=[]
for i in range(0,299):    
    img_sample_cat = stylegan_psi1_test[i]
    basewidth = 512
    img_sample_cat = Image.open(img_sample_cat)
    wpercent = (basewidth/float(img_sample_cat.size[0]))
    hsize = 512
    img_sample_cat = img_sample_cat.resize((basewidth,hsize), Image.ANTIALIAS)
    
    sample_res=prnu.extract_single(np.array(img_sample_cat)) #extract sıngle pıcture resıduals
    sample_res = np.stack(sample_res, 0)
    from numpy import zeros, newaxis
    sample_res = sample_res[newaxis,:, :]
    pred=pd.DataFrame(prnu.aligned_cc(sample_res,c)['ncc']) #ncc between test resıduals and normal pıctures fıngerprınts
    if pred.values >0.001:  #threshold
        predıctıon_values.append(0)
        corr.append(pred.values)
    else:
        predıctıon_values.append(1)
        corr.append(pred.values)
predıctıon_values=pd.DataFrame(predıctıon_values)
predıctıon_values["label"]=1
predıctıon_values.columns=["Pred","label"]
predıctıon_values.head()
from sklearn.metrics import accuracy_score
print("Style GAN psi1 Accuracy Score:",accuracy_score(predıctıon_values.label,predıctıon_values.Pred))
print("Style GAN psi1 correlation mean: ", np.mean(corr))

Style GAN psi1 Accuracy Score: 0.6555183946488294
Style GAN psi1 correlation mean:  7.933077e-05

predıctıon_values=[]
corr=[]
for i in range(0,299):    
    img_sample_cat = stylegan2_psi05_test[i]
    basewidth = 512
    img_sample_cat = Image.open(img_sample_cat)
    wpercent = (basewidth/float(img_sample_cat.size[0]))
    hsize = 512
    img_sample_cat = img_sample_cat.resize((basewidth,hsize), Image.ANTIALIAS)
    
    sample_res=prnu.extract_single(np.array(img_sample_cat)) #extract sıngle pıcture resıduals
    sample_res = np.stack(sample_res, 0)
    from numpy import zeros, newaxis
    sample_res = sample_res[newaxis,:, :]
    pred=pd.DataFrame(prnu.aligned_cc(sample_res,c)['ncc']) #ncc between test resıduals and normal pıctures fıngerprınts
    if pred.values >0.001:  #threshold
        predıctıon_values.append(0)
        corr.append(pred.values)
    else:
        predıctıon_values.append(1)
        corr.append(pred.values)
predıctıon_values=pd.DataFrame(predıctıon_values)
predıctıon_values["label"]=1
predıctıon_values.columns=["Pred","label"]
predıctıon_values.head()
from sklearn.metrics import accuracy_score
print("Style GAN2 psi05 Accuracy Score:",accuracy_score(predıctıon_values.label,predıctıon_values.Pred))
print("Style GAN2 psi05 correlation mean: ", np.mean(corr))

Style GAN2 psi05 Accuracy Score: 0.5451505016722408
Style GAN2 psi05 correlation mean:  0.00067534804

predıctıon_values=[]
corr=[]
for i in range(0,299):    
    img_sample_cat = stylegan2_psi1_test[i]
    basewidth = 512
    img_sample_cat = Image.open(img_sample_cat)
    wpercent = (basewidth/float(img_sample_cat.size[0]))
    hsize = 512
    img_sample_cat = img_sample_cat.resize((basewidth,hsize), Image.ANTIALIAS)
    
    sample_res=prnu.extract_single(np.array(img_sample_cat)) #extract sıngle pıcture resıduals
    sample_res = np.stack(sample_res, 0)
    from numpy import zeros, newaxis
    sample_res = sample_res[newaxis,:, :]
    pred=pd.DataFrame(prnu.aligned_cc(sample_res,c)['ncc']) #ncc between test resıduals and normal pıctures fıngerprınts
    if pred.values >0.001:  #threshold
        predıctıon_values.append(0)
        corr.append(pred.values)
    else:
        predıctıon_values.append(1)
        corr.append(pred.values)
predıctıon_values=pd.DataFrame(predıctıon_values)
predıctıon_values["label"]=1
predıctıon_values.columns=["Pred","label"]
predıctıon_values.head()
from sklearn.metrics import accuracy_score
print("Style GAN2 psi1 Accuracy Score:",accuracy_score(predıctıon_values.label,predıctıon_values.Pred))
print("Style GAN2 psi1 correlation mean: ", np.mean(corr))

Style GAN2 psi1 Accuracy Score: 0.5953177257525084
Style GAN2 psi1 correlation mean:  0.00025719954

Test with threshold = 0.005

nat_dirlist_test

array(['test\\data\\test\\normal\\05304.png',
       'test\\data\\test\\normal\\05305.png',
       'test\\data\\test\\normal\\05306.png',
       'test\\data\\test\\normal\\05307.png',
       'test\\data\\test\\normal\\05308.png',
       'test\\data\\test\\normal\\05309.png',
       'test\\data\\test\\normal\\05310.png',
       'test\\data\\test\\normal\\05311.png',
       'test\\data\\test\\normal\\05312.png',
       'test\\data\\test\\normal\\05313.png',
       'test\\data\\test\\normal\\05314.png',
       'test\\data\\test\\normal\\05315.png',
       'test\\data\\test\\normal\\05316.png',
       'test\\data\\test\\normal\\05317.png',
       'test\\data\\test\\normal\\05318.png',
       'test\\data\\test\\normal\\05319.png',
       'test\\data\\test\\normal\\05320.png',
       'test\\data\\test\\normal\\05321.png',
       'test\\data\\test\\normal\\05322.png',
       'test\\data\\test\\normal\\05323.png',
       'test\\data\\test\\normal\\05324.png',
       'test\\data\\test\\normal\\05325.png',
       'test\\data\\test\\normal\\05326.png',
       'test\\data\\test\\normal\\05327.png',
       'test\\data\\test\\normal\\05328.png',
       'test\\data\\test\\normal\\05329.png',
       'test\\data\\test\\normal\\05330.png',
       'test\\data\\test\\normal\\05331.png',
       'test\\data\\test\\normal\\05332.png',
       'test\\data\\test\\normal\\05333.png',
       'test\\data\\test\\normal\\05334.png',
       'test\\data\\test\\normal\\05335.png',
       'test\\data\\test\\normal\\05336.png',
       'test\\data\\test\\normal\\05337.png',
       'test\\data\\test\\normal\\05338.png',
       'test\\data\\test\\normal\\05339.png',
       'test\\data\\test\\normal\\05340.png',
       'test\\data\\test\\normal\\05341.png',
       'test\\data\\test\\normal\\05342.png',
       'test\\data\\test\\normal\\05343.png',
       'test\\data\\test\\normal\\05344.png',
       'test\\data\\test\\normal\\05345.png',
       'test\\data\\test\\normal\\05346.png',
       'test\\data\\test\\normal\\05347.png',
       'test\\data\\test\\normal\\05348.png',
       'test\\data\\test\\normal\\05349.png',
       'test\\data\\test\\normal\\05350.png',
       'test\\data\\test\\normal\\05351.png',
       'test\\data\\test\\normal\\05352.png',
       'test\\data\\test\\normal\\05353.png',
       'test\\data\\test\\normal\\05354.png',
       'test\\data\\test\\normal\\05355.png',
       'test\\data\\test\\normal\\05356.png',
       'test\\data\\test\\normal\\05357.png',
       'test\\data\\test\\normal\\05358.png',
       'test\\data\\test\\normal\\05359.png',
       'test\\data\\test\\normal\\05360.png',
       'test\\data\\test\\normal\\05361.png',
       'test\\data\\test\\normal\\05362.png',
       'test\\data\\test\\normal\\05363.png',
       'test\\data\\test\\normal\\05364.png',
       'test\\data\\test\\normal\\05365.png',
       'test\\data\\test\\normal\\05366.png',
       'test\\data\\test\\normal\\05367.png',
       'test\\data\\test\\normal\\05368.png',
       'test\\data\\test\\normal\\05369.png',
       'test\\data\\test\\normal\\05370.png',
       'test\\data\\test\\normal\\05371.png',
       'test\\data\\test\\normal\\05372.png',
       'test\\data\\test\\normal\\05373.png',
       'test\\data\\test\\normal\\05374.png',
       'test\\data\\test\\normal\\05375.png',
       'test\\data\\test\\normal\\05376.png',
       'test\\data\\test\\normal\\05377.png',
       'test\\data\\test\\normal\\05378.png',
       'test\\data\\test\\normal\\05379.png',
       'test\\data\\test\\normal\\05380.png',
       'test\\data\\test\\normal\\05381.png',
       'test\\data\\test\\normal\\05382.png',
       'test\\data\\test\\normal\\05383.png',
       'test\\data\\test\\normal\\05384.png',
       'test\\data\\test\\normal\\05385.png',
       'test\\data\\test\\normal\\05386.png',
       'test\\data\\test\\normal\\05387.png',
       'test\\data\\test\\normal\\05388.png',
       'test\\data\\test\\normal\\05389.png',
       'test\\data\\test\\normal\\05390.png',
       'test\\data\\test\\normal\\05391.png',
       'test\\data\\test\\normal\\05392.png',
       'test\\data\\test\\normal\\05393.png',
       'test\\data\\test\\normal\\05394.png',
       'test\\data\\test\\normal\\05395.png',
       'test\\data\\test\\normal\\05396.png',
       'test\\data\\test\\normal\\05397.png',
       'test\\data\\test\\normal\\05398.png',
       'test\\data\\test\\normal\\05399.png',
       'test\\data\\test\\normal\\05400.png',
       'test\\data\\test\\normal\\05401.png',
       'test\\data\\test\\normal\\05402.png',
       'test\\data\\test\\normal\\05403.png',
       'test\\data\\test\\normal\\05404.png',
       'test\\data\\test\\normal\\05405.png',
       'test\\data\\test\\normal\\05406.png',
       'test\\data\\test\\normal\\05407.png',
       'test\\data\\test\\normal\\05408.png',
       'test\\data\\test\\normal\\05409.png',
       'test\\data\\test\\normal\\05410.png',
       'test\\data\\test\\normal\\05411.png',
       'test\\data\\test\\normal\\05412.png',
       'test\\data\\test\\normal\\05413.png',
       'test\\data\\test\\normal\\05414.png',
       'test\\data\\test\\normal\\05415.png',
       'test\\data\\test\\normal\\05416.png',
       'test\\data\\test\\normal\\05417.png',
       'test\\data\\test\\normal\\05418.png',
       'test\\data\\test\\normal\\05419.png',
       'test\\data\\test\\normal\\05420.png',
       'test\\data\\test\\normal\\05421.png',
       'test\\data\\test\\normal\\05422.png',
       'test\\data\\test\\normal\\05423.png',
       'test\\data\\test\\normal\\05424.png',
       'test\\data\\test\\normal\\05425.png',
       'test\\data\\test\\normal\\05426.png',
       'test\\data\\test\\normal\\05427.png',
       'test\\data\\test\\normal\\05428.png',
       'test\\data\\test\\normal\\05429.png',
       'test\\data\\test\\normal\\05430.png',
       'test\\data\\test\\normal\\05431.png',
       'test\\data\\test\\normal\\05432.png',
       'test\\data\\test\\normal\\05433.png',
       'test\\data\\test\\normal\\05434.png',
       'test\\data\\test\\normal\\05435.png',
       'test\\data\\test\\normal\\05436.png',
       'test\\data\\test\\normal\\05437.png',
       'test\\data\\test\\normal\\05438.png',
       'test\\data\\test\\normal\\05439.png',
       'test\\data\\test\\normal\\05440.png',
       'test\\data\\test\\normal\\05441.png',
       'test\\data\\test\\normal\\05442.png',
       'test\\data\\test\\normal\\05443.png',
       'test\\data\\test\\normal\\05444.png',
       'test\\data\\test\\normal\\05445.png',
       'test\\data\\test\\normal\\05446.png',
       'test\\data\\test\\normal\\05447.png',
       'test\\data\\test\\normal\\05448.png',
       'test\\data\\test\\normal\\05449.png',
       'test\\data\\test\\normal\\05450.png',
       'test\\data\\test\\normal\\05451.png',
       'test\\data\\test\\normal\\05452.png',
       'test\\data\\test\\normal\\05453.png',
       'test\\data\\test\\normal\\05454.png',
       'test\\data\\test\\normal\\05455.png',
       'test\\data\\test\\normal\\05456.png',
       'test\\data\\test\\normal\\05457.png',
       'test\\data\\test\\normal\\05458.png',
       'test\\data\\test\\normal\\05459.png',
       'test\\data\\test\\normal\\05460.png',
       'test\\data\\test\\normal\\05461.png',
       'test\\data\\test\\normal\\05462.png',
       'test\\data\\test\\normal\\05463.png',
       'test\\data\\test\\normal\\05464.png',
       'test\\data\\test\\normal\\05465.png',
       'test\\data\\test\\normal\\05466.png',
       'test\\data\\test\\normal\\05467.png',
       'test\\data\\test\\normal\\05468.png',
       'test\\data\\test\\normal\\05469.png',
       'test\\data\\test\\normal\\05470.png',
       'test\\data\\test\\normal\\05471.png',
       'test\\data\\test\\normal\\05472.png',
       'test\\data\\test\\normal\\05473.png',
       'test\\data\\test\\normal\\05474.png',
       'test\\data\\test\\normal\\05475.png',
       'test\\data\\test\\normal\\05476.png',
       'test\\data\\test\\normal\\05477.png',
       'test\\data\\test\\normal\\05478.png',
       'test\\data\\test\\normal\\05479.png',
       'test\\data\\test\\normal\\05480.png',
       'test\\data\\test\\normal\\05481.png',
       'test\\data\\test\\normal\\05482.png',
       'test\\data\\test\\normal\\05483.png',
       'test\\data\\test\\normal\\05484.png',
       'test\\data\\test\\normal\\05485.png',
       'test\\data\\test\\normal\\05486.png',
       'test\\data\\test\\normal\\05487.png',
       'test\\data\\test\\normal\\05488.png',
       'test\\data\\test\\normal\\05489.png',
       'test\\data\\test\\normal\\05490.png',
       'test\\data\\test\\normal\\05491.png',
       'test\\data\\test\\normal\\05492.png',
       'test\\data\\test\\normal\\05493.png',
       'test\\data\\test\\normal\\05494.png',
       'test\\data\\test\\normal\\05495.png',
       'test\\data\\test\\normal\\05496.png',
       'test\\data\\test\\normal\\05497.png',
       'test\\data\\test\\normal\\05498.png',
       'test\\data\\test\\normal\\05499.png',
       'test\\data\\test\\normal\\05500.png',
       'test\\data\\test\\normal\\05501.png',
       'test\\data\\test\\normal\\05502.png',
       'test\\data\\test\\normal\\05503.png',
       'test\\data\\test\\normal\\05504.png',
       'test\\data\\test\\normal\\05505.png',
       'test\\data\\test\\normal\\05506.png',
       'test\\data\\test\\normal\\05507.png',
       'test\\data\\test\\normal\\05508.png',
       'test\\data\\test\\normal\\05509.png',
       'test\\data\\test\\normal\\05510.png',
       'test\\data\\test\\normal\\05511.png',
       'test\\data\\test\\normal\\05512.png',
       'test\\data\\test\\normal\\05513.png',
       'test\\data\\test\\normal\\05514.png',
       'test\\data\\test\\normal\\05515.png',
       'test\\data\\test\\normal\\05516.png',
       'test\\data\\test\\normal\\05517.png',
       'test\\data\\test\\normal\\05518.png',
       'test\\data\\test\\normal\\05519.png',
       'test\\data\\test\\normal\\05520.png',
       'test\\data\\test\\normal\\05521.png',
       'test\\data\\test\\normal\\05522.png',
       'test\\data\\test\\normal\\05523.png',
       'test\\data\\test\\normal\\05524.png',
       'test\\data\\test\\normal\\05525.png',
       'test\\data\\test\\normal\\05526.png',
       'test\\data\\test\\normal\\05527.png',
       'test\\data\\test\\normal\\05528.png',
       'test\\data\\test\\normal\\05529.png',
       'test\\data\\test\\normal\\05530.png',
       'test\\data\\test\\normal\\05531.png',
       'test\\data\\test\\normal\\05532.png',
       'test\\data\\test\\normal\\05533.png',
       'test\\data\\test\\normal\\05534.png',
       'test\\data\\test\\normal\\05535.png',
       'test\\data\\test\\normal\\05536.png',
       'test\\data\\test\\normal\\05537.png',
       'test\\data\\test\\normal\\05538.png',
       'test\\data\\test\\normal\\05539.png',
       'test\\data\\test\\normal\\05540.png',
       'test\\data\\test\\normal\\05541.png',
       'test\\data\\test\\normal\\05542.png',
       'test\\data\\test\\normal\\05543.png',
       'test\\data\\test\\normal\\05544.png',
       'test\\data\\test\\normal\\05545.png',
       'test\\data\\test\\normal\\05546.png',
       'test\\data\\test\\normal\\05547.png',
       'test\\data\\test\\normal\\05548.png',
       'test\\data\\test\\normal\\05549.png',
       'test\\data\\test\\normal\\05550.png',
       'test\\data\\test\\normal\\05551.png',
       'test\\data\\test\\normal\\05552.png',
       'test\\data\\test\\normal\\05553.png',
       'test\\data\\test\\normal\\05554.png',
       'test\\data\\test\\normal\\05555.png',
       'test\\data\\test\\normal\\05556.png',
       'test\\data\\test\\normal\\05557.png',
       'test\\data\\test\\normal\\05558.png',
       'test\\data\\test\\normal\\05559.png',
       'test\\data\\test\\normal\\05560.png',
       'test\\data\\test\\normal\\05561.png',
       'test\\data\\test\\normal\\05562.png',
       'test\\data\\test\\normal\\05563.png',
       'test\\data\\test\\normal\\05564.png',
       'test\\data\\test\\normal\\05565.png',
       'test\\data\\test\\normal\\05566.png',
       'test\\data\\test\\normal\\05567.png',
       'test\\data\\test\\normal\\05568.png',
       'test\\data\\test\\normal\\05569.png',
       'test\\data\\test\\normal\\05570.png',
       'test\\data\\test\\normal\\05571.png',
       'test\\data\\test\\normal\\05572.png',
       'test\\data\\test\\normal\\05573.png',
       'test\\data\\test\\normal\\05574.png',
       'test\\data\\test\\normal\\05575.png',
       'test\\data\\test\\normal\\05576.png',
       'test\\data\\test\\normal\\05577.png',
       'test\\data\\test\\normal\\05578.png',
       'test\\data\\test\\normal\\05579.png',
       'test\\data\\test\\normal\\05580.png',
       'test\\data\\test\\normal\\05581.png',
       'test\\data\\test\\normal\\05582.png',
       'test\\data\\test\\normal\\05583.png',
       'test\\data\\test\\normal\\05584.png',
       'test\\data\\test\\normal\\05585.png',
       'test\\data\\test\\normal\\05586.png',
       'test\\data\\test\\normal\\05587.png',
       'test\\data\\test\\normal\\05588.png',
       'test\\data\\test\\normal\\05589.png',
       'test\\data\\test\\normal\\05590.png',
       'test\\data\\test\\normal\\05591.png',
       'test\\data\\test\\normal\\05592.png',
       'test\\data\\test\\normal\\05593.png',
       'test\\data\\test\\normal\\05594.png',
       'test\\data\\test\\normal\\05595.png',
       'test\\data\\test\\normal\\05596.png',
       'test\\data\\test\\normal\\05597.png',
       'test\\data\\test\\normal\\05598.png',
       'test\\data\\test\\normal\\05599.png',
       'test\\data\\test\\normal\\05600.png',
       'test\\data\\test\\normal\\05601.png',
       'test\\data\\test\\normal\\05602.png',
       'test\\data\\test\\normal\\05603.png'], dtype='<U31')

print('Computing residuals')

imgs = []
for img_path in nat_dirlist_test:
    imgs += [prnu.cut_ctr(np.asarray(Image.open(img_path)), (512, 512, 3))]

pool = Pool(cpu_count())
w_test = pool.map(prnu.extract_single, imgs)
pool.close()

w_test = np.stack(w_test, 0)

Computing residuals

predıctıon_values=[]
corr=[]
for i in range(0,299):    
    img_sample_cat = nat_dirlist_test[i]
    basewidth = 512
    img_sample_cat = Image.open(img_sample_cat)
    wpercent = (basewidth/float(img_sample_cat.size[0]))
    hsize = 512
    img_sample_cat = img_sample_cat.resize((basewidth,hsize), Image.ANTIALIAS)
    
    sample_res=prnu.extract_single(np.array(img_sample_cat)) #extract sıngle pıcture resıduals
    sample_res = np.stack(sample_res, 0)
    from numpy import zeros, newaxis
    sample_res = sample_res[newaxis,:, :]
    pred=pd.DataFrame(prnu.aligned_cc(c,sample_res)['ncc']) #ncc between test resıduals and normal pıctures fıngerprınts
    if pred.values >0.005:  #threshold
        predıctıon_values.append(0)
        corr.append(pred.values)
    else:
        predıctıon_values.append(1)
        corr.append(pred.values)
predıctıon_values=pd.DataFrame(predıctıon_values)
predıctıon_values["label"]=1
predıctıon_values.columns=["Pred","label"]
predıctıon_values.head()
from sklearn.metrics import accuracy_score
print("Normal Accuracy Score:",accuracy_score(predıctıon_values.label,predıctıon_values.Pred))

Normal Accuracy Score: 0.9732441471571907

predıctıon_values=[]
corr=[]
for i in range(0,299):    
    img_sample_cat = stylegan_psi05_test[i]
    basewidth = 512
    img_sample_cat = Image.open(img_sample_cat)
    wpercent = (basewidth/float(img_sample_cat.size[0]))
    hsize = 512
    img_sample_cat = img_sample_cat.resize((basewidth,hsize), Image.ANTIALIAS)
    
    sample_res=prnu.extract_single(np.array(img_sample_cat)) #extract sıngle pıcture resıduals
    sample_res = np.stack(sample_res, 0)
    from numpy import zeros, newaxis
    sample_res = sample_res[newaxis,:, :]
    pred=pd.DataFrame(prnu.aligned_cc(sample_res,c)['ncc']) #ncc between test resıduals and normal pıctures fıngerprınts
    if pred.values >0.005:  #threshold
        predıctıon_values.append(0)
        corr.append(pred.values)
    else:
        predıctıon_values.append(1)
        corr.append(pred.values)
predıctıon_values=pd.DataFrame(predıctıon_values)
predıctıon_values["label"]=1
predıctıon_values.columns=["Pred","label"]
predıctıon_values.head()
from sklearn.metrics import accuracy_score
print("Style GAN psi05 Accuracy Score:",accuracy_score(predıctıon_values.label,predıctıon_values.Pred))

Style GAN psi05 Accuracy Score: 0.9698996655518395

predıctıon_values=[]
corr=[]
for i in range(0,299):    
    img_sample_cat = stylegan_psi1_test[i]
    basewidth = 512
    img_sample_cat = Image.open(img_sample_cat)
    wpercent = (basewidth/float(img_sample_cat.size[0]))
    hsize = 512
    img_sample_cat = img_sample_cat.resize((basewidth,hsize), Image.ANTIALIAS)
    
    sample_res=prnu.extract_single(np.array(img_sample_cat)) #extract sıngle pıcture resıduals
    sample_res = np.stack(sample_res, 0)
    from numpy import zeros, newaxis
    sample_res = sample_res[newaxis,:, :]
    pred=pd.DataFrame(prnu.aligned_cc(sample_res,c)['ncc']) #ncc between test resıduals and normal pıctures fıngerprınts
    if pred.values >0.005:  #threshold
        predıctıon_values.append(0)
        corr.append(pred.values)
    else:
        predıctıon_values.append(1)
        corr.append(pred.values)
predıctıon_values=pd.DataFrame(predıctıon_values)
predıctıon_values["label"]=1
predıctıon_values.columns=["Pred","label"]
predıctıon_values.head()
from sklearn.metrics import accuracy_score
print("Style GAN psi1 Accuracy Score:",accuracy_score(predıctıon_values.label,predıctıon_values.Pred))

Style GAN psi1 Accuracy Score: 0.9732441471571907

predıctıon_values=[]
corr=[]
for i in range(0,299):    
    img_sample_cat = stylegan2_psi05_test[i]
    basewidth = 512
    img_sample_cat = Image.open(img_sample_cat)
    wpercent = (basewidth/float(img_sample_cat.size[0]))
    hsize = 512
    img_sample_cat = img_sample_cat.resize((basewidth,hsize), Image.ANTIALIAS)
    
    sample_res=prnu.extract_single(np.array(img_sample_cat)) #extract sıngle pıcture resıduals
    sample_res = np.stack(sample_res, 0)
    from numpy import zeros, newaxis
    sample_res = sample_res[newaxis,:, :]
    pred=pd.DataFrame(prnu.aligned_cc(sample_res,c)['ncc']) #ncc between test resıduals and normal pıctures fıngerprınts
    if pred.values >0.005:  #threshold
        predıctıon_values.append(0)
        corr.append(pred.values)
    else:
        predıctıon_values.append(1)
        corr.append(pred.values)
predıctıon_values=pd.DataFrame(predıctıon_values)
predıctıon_values["label"]=1
predıctıon_values.columns=["Pred","label"]
predıctıon_values.head()
from sklearn.metrics import accuracy_score
print("Style GAN2 psi05 Accuracy Score:",accuracy_score(predıctıon_values.label,predıctıon_values.Pred))

Style GAN2 psi05 Accuracy Score: 0.9531772575250836

predıctıon_values=[]
corr=[]
for i in range(0,299):    
    img_sample_cat = stylegan2_psi1_test[i]
    basewidth = 512
    img_sample_cat = Image.open(img_sample_cat)
    wpercent = (basewidth/float(img_sample_cat.size[0]))
    hsize = 512
    img_sample_cat = img_sample_cat.resize((basewidth,hsize), Image.ANTIALIAS)
    
    sample_res=prnu.extract_single(np.array(img_sample_cat)) #extract sıngle pıcture resıduals
    sample_res = np.stack(sample_res, 0)
    from numpy import zeros, newaxis
    sample_res = sample_res[newaxis,:, :]
    pred=pd.DataFrame(prnu.aligned_cc(sample_res,c)['ncc']) #ncc between test resıduals and normal pıctures fıngerprınts
    if pred.values >0.005:  #threshold
        predıctıon_values.append(0)
        corr.append(pred.values)
    else:
        predıctıon_values.append(1)
        corr.append(pred.values)
predıctıon_values=pd.DataFrame(predıctıon_values)
predıctıon_values["label"]=1
predıctıon_values.columns=["Pred","label"]
predıctıon_values.head()
from sklearn.metrics import accuracy_score
print("Style GAN2 psi1 Accuracy Score:",accuracy_score(predıctıon_values.label,predıctıon_values.Pred))

Style GAN2 psi1 Accuracy Score: 0.9866220735785953